How to use this presentation

• This presentation should not be distributed in physical or electronic form and the
contents should never be used in public domain (e.g. blogs, twitter, papers, etc.)
• The content of this presentation can be presented to customers.
• Any projected performance conversations should first be modeled in the
VxRail Sizing Tool ([Link]) and then shared with customers.

Copyright © Dell Inc. All Rights Reserved.

Internal Use - Confidential
Dell VxRail
Performance &
Sizing Guide
July 2025
 Table of contents
Sizing Methodology and Using the VxRail Sizing Tool • VMware Subscription Licensing considerations
(UPDATED April 2025!)
• Requirements Defined
• Performance of VxRail 16th
• Using the VxRail Sizing Tool
Generation with AMD EPYC Gen 4 on vSAN ESA and O
SA
• VxRail platforms: Performance of VD-4000 on vSAN OS
Sizing Guidance and Testing Data A & ESA 8
• Importance of cache drives with vSAN OSA
• Creating a vSAN OSA configuration • VxRail platforms: Performance of 16G Sapphire Rapids
Capacity drives & disk groups • AI Benchmarking with VxRail
• Cache & Capacity Drive Evaluation • VD-4000 Performance and Power Efficiency
• Networking considerations for vSAN OSA • General processor considerations
• Networking 100GbE for vSAN ESA • VxRail with vSAN 8.0 ESA
• vSAN Over RDMA • VxRail ESA TPC-DS Analysis
• Memory considerations
• CPU Considerations

Internal Use - Confidential 3 Copyright © Dell Inc. All Rights Reserved.

 Requirements
Defined

Internal Use - Confidential 4 Copyright © Dell Inc. All Rights Reserved.

 The Sizing Challenge
Sizing VxRail is finding the balance between:
• Functional Requirements:
o HA capabilities
o All Flash/Hybrid

Cost $
o Form factor
Requirement

Performance
Functional

• Performance:
o IO performance
o Storage capacity
o CPU performance
o Network performance
o Memory performance

• Cost of the Solution

o VxRail H/W configuration
o Licensing etc
Balanced Configuration
• The Sizing Tool will help:
o Finding the most balanced configuration
Sizing your clusters is both a science and an art form, with the o Achieve all the requirements specified
quality of the sizing highly on how depending on how accurately o But you will still need to refine the output
we define the requirements.
Internal Use - Confidential 5 Copyright © Dell Inc. All Rights Reserved.
 VxRail Sizing Workflow Overview

Requirements Systems Evaluation &

Quote
Definition Sizing Customization

Internal Use - Confidential 6 Copyright © Dell Inc. All Rights Reserved.

 Designing and sizing your VxRail system
One of the core objectives of a sizing exercise is to determine the pools of resources
needed to satisfy workload demands and their SLOs, at an optimal cost

Capacity Expectation Cost ($)

“How much I need” “What do I get?” “How much does it cost?”

”
Service level expectations

The quality of the sizing depends highly on how accurately we define the requirements

Internal Use - Confidential 7 Copyright © Dell Inc. All Rights Reserved.

 The accuracy of the requirements vary according to
the use case

Technology Refresh New Environments Greenfield

(VM consolidation, migration from (frame of reference) (no references)
legacy HW, Test & Dev, etc.)

• In this case we have the information • We don’t have specific information of • Little information is available about the
of the applications that we are sizing what workloads will be running, but we workloads that will run in the new
for can use an existing application system
• It’s extremely important to use the environment as a frame of reference • Working in collaboration with the
data collection and summarization • Preferably use LiveOptics to customer, use a set of “Reference
tools, such as LiveOptics to identify characterize the workloads to be used Workloads” to define a system
the requirements when working on as reference workload profile and provide
Tech Refresh and/or consolidation • Note that sizing based on simple reference
use cases
reproduction of existing HW
characteristics can be done, but we
may miss opportunities to optimize the
configuration.

Internal Use - Confidential 8 Copyright © Dell Inc. All Rights Reserved.

 Defining workload requirements

Technology refresh use cases

• use data collected with tools such as Dell LiveOptics

When “real life” data is not available

• Leverage reference workloads in the VxRail Sizing Tool
• Reference loads are synthetic workloads that represent real life
workloads
o Databases (DBMS)
o online transactions (OLTP)
o Public benchmarks (TPC-C, TPC-H, etc.)

VxRail Sizing Tool

I need a system that can • Can import data retrieved from LiveOptics analysis
support 1000 x VMs. Each one with
• Data is summarized at the VM level and optimized by the sizing
• 250GB of storage capacity
• 1200 MHz of CPU capacity tool
• 4 GB of Memory capacity
Work together with your SPS rep on sizing/modeling
We estimate that each VM will be running exercises
~ 200 IOPS of a TPC-C workload

Internal Use - Confidential 9 Copyright © Dell Inc. All Rights Reserved.

 Other configuration requirements
E Series

Specific application needs Datacenter footprint Local protection using

can affect the processor • When physical space is snapshots or RP4VM, has
model decision constraint, a customer may additional capacity
• VDI will benefit from CPUs prefer a 1U model like the E- requirements
with higher core count series if drive and PCIe
limitations are acceptable
• Single threaded apps benefit
from CPUs with higher clock
speed

Internal Use - Confidential 10 Copyright © Dell Inc. All Rights Reserved.

 Cluster size considerations

• Based on vSAN Cluster Design: vSAN Planning and Deployment

• Cluster sizing for customer needs is critical, there is no single best answer for node
counts
• A general rule of guidance is that a node takes approximately 1.5 hours to update,
update times can approach 24 hours larger clusters.
• Customers need to make sure to consider their operational experience during
cluster sizing and planning.
• The tradeoffs that follow are a non-exhaustive list and are the most common ones.

Internal Use - Confidential 11 Copyright © Dell Inc. All Rights Reserved.

 Cluster size tradeoff considerations

Upgrade Time Prescriptive Cluster Design Automation @ Scale

Advantage: Small Clusters Advantage: Small clusters Advantage: Small clusters

Smaller clusters complete upgrades Using more smaller clusters allows A greater number of smaller clusters
more quickly. A 5-node cluster can be administrators to assign cluster requires more actions to manage all
updated in approximately 8 hours. resources and services on an as- environments. This makes numerous
needed basis and with narrower sets small cluster environments excellent
of data services in each cluster. candidates for automation!

Larger clusters take more time to Using larger general-purpose clusters Consolidating nodes into larger
complete an upgrade.16 nodes could puts more resources in fewer fault clusters and resource pools reduces
be upgraded in approximately 24 domains. More policies may be the opportunities to leverage
hours. needed to serve all workloads automation.
consolidated.

Internal Use - Confidential 12 Copyright © Dell Inc. All Rights Reserved.

 Cluster size tradeoff considerations

Simplicity Buffer Resource Efficiency Resiliency

Advantage: Large Clusters Advantage: Large Clusters
Advantage: Large Clusters
Deploying more numerous smaller Small clusters use a higher Small clusters can be limited in their
clusters creates a more complex percentage of resources to achieve storage and protection policies, and
environment of tracking and similar protection as large clusters. have fewer resources for rebuilds
management. N+1 protection comes at a 20% cost and resync operations.
for a 5-node cluster.

Larger general-purpose clusters are Large clusters are more efficient! Large clusters can more easily
easier to plan and scale. Just add Achieve N+1 protection in a 16-node achieve protection goals as
nodes as needed to grow the cluster at a 6% resource cost. resiliency options can be satisfied
environment! with more nodes, and the can
recover more quickly from rebuild
and resync operations

Internal Use - Confidential 13 Copyright © Dell Inc. All Rights Reserved.

 Using the VxRail Sizing Tool
Internal Use - Confidential 14 Copyright © Dell Inc. All Rights Reserved.
 Up
da

VxRail Sizing Tool usage tutorial te

d!

Find detailed information on how to use the VxRail Sizing tool from the tool’s help menu

• Access the VxRail Sizing Tool Help Menu Tutorial here

Internal Use - Confidential 15 Copyright © Dell Inc. All Rights Reserved.

 Importance of cache
drives with vSAN
OSA

Internal Use - Confidential 16 Copyright © Dell Inc. All Rights Reserved.

 Evaluate benefit of different cache drive types
• Different cache drive types are available in the Sizing tool

Mixed use SAS Write intensive SAS NVMe

• Overall performance will be affected by the drive types chosen for the cache or capacity tier

• It is important to know beforehand the type of workload that is going to be deployed on the cluster and
also the drive specs.

Internal Use - Confidential 17 Copyright © Dell Inc. All Rights Reserved.

 Cache Drive Types Quick Facts
Main Observations

Perf differences For small block For larger block

between workloads (4k/8k) workloads there is a
SAS/NVMe starts SAS/NVMe clear performance
above 50% max performance is differentiation between
IOPS very similar even SAS/NVMe
at higher IOPS
workloads

Internal Use - Confidential 18 Copyright © Dell Inc. All Rights Reserved.

 Cache capacity is critical

Configure for a 10% cache-to-capacity

ratio for Hybrid systems

A larger cache size has a direct benefit to the

performance of Hybrid systems
For medium and heavy write For read intensive and generic
workloads (>30%) increase workloads the total cache
cache capacity per host capacity/host can be smaller

Internal Use - Confidential 19 Copyright © Dell Inc. All Rights Reserved.

 vSAN OSA Large Write
Buffer benefits 8.0.000
and later
Internal Use - Confidential 20 Copyright © Dell Inc. All Rights Reserved.
 VxRail test environments
Large write buffer test beds

VXRAIL 7.0.350 VXRAIL 8.0.000

• 4 x VxRail P570F • 4 x VxRail P570F

• 2 x Intel(R) Xeon(R) Gold • 2 x Intel(R) Xeon(R) Gold
6226R CPU @ 2.9 GHz 6226R CPU @ 2.9 GHz
 32 Cores per node  32 Cores per node

• 768 GB RAM • 768 GB RAM

• Broadcom Adv. 25 GbE • Broadcom Adv. 25 GbE
• 2 Disk Groups per host: • 2 Disk Groups per host:
 1 x Cache SAS  1 x Cache SAS
"PowerEdge Agnostic“ "PowerEdge Agnostic“
1.6 TB 1.6 TB
 3 x Capacity SAS  3 x Capacity SAS Kioxia
Kioxia PM5 3.8 TB PM5 3.8 TB
• Write Buffers = 600 GB • Write Buffers ~ 1.6 TB Note: You must enable large cache by
following the procedures in Solve
These 2 systems were exactly the same:
We just deployed different code levels & WB size

Internal Use - Confidential 21 Copyright © Dell Inc. All Rights Reserved.

 Verifying large write buffers

 From an ESXi host, can run a shell command using vsish. For example, the
following can be used from the master host:

esxcli vsan storage list | grep "VSAN Disk Group UUID" | awk -F ': ' '{print
$2}'| uniq | xargs -I {} vsish -e get /vmkModules/lsom/disks/{}/info | grep
"Write Buffer Size" |awk -F':' '{print $1": "$2/1000/1000/1000/1000" TB
(Terabyte)"}’
 Will get 1 output line per host like the following for a 4-node cluster:

Internal Use - Confidential 22 Copyright © Dell Inc. All Rights Reserved.

 Resync Test – OLTP 16KB at 50% of max IOPS
2-hour steady state test, RAID1FTT1, SE = none, 1 Disk Group removed at 30 mins

VxRail P570F, 4 Nodes, SAS 1.6 TB Cache, Small WB VxRail P570F, 4 Nodes, SAS 1.6 TB Cache, Large WB
SAS 3.8 TB Capacity, OLTP 16K, Moderate IOPS (170K) SAS 3.8 TB Capacity, OLTP 16K, Moderate IOPS (170K)
3.5 3.5
Resync time: 75:40 Resync time: 46:38
Data migrated: 3.2 TB 3.0 Data migrated: 3.2 TB
3.0
Migration thruput: 743.7 Migration thruput: 1,212,6
MB/s MB/s
2.5 2.5

Response Time (ms)

Response Time (ms)

2.0 2.0 - 38% resync time

+ 63% throughput
1.5 1.5

1.0 1.0
1 Disk group
1 Disk group
0.5 removed
0.5 removed

OLTP_16K_50pct, RAID1, no SE, WB 600 GB OLTP_16K_50pct, RAID1, no SE, WB 1.6 TB

0.0 0.0
0 10 20 30 40 50 60 70 80 90 100 110 120 0 10 20 30 40 50 60 70 80 90 100 110 120

Elapsed Time - Minutes Elapsed Time - Minutes

Internal Use - Confidential 23 Copyright © Dell Inc. All Rights Reserved.

 Key Takeaways – Large write buffers

• Biggest improvements were in resync tests, as total resync time dropped 38% due
to 63% increase in resync throughput
• Advise customers with larger cache drive to enable this in vSAN 8.0 follow the
Solve procedures

• Note 1: For characterization tests, some very modest improvements were noted, especially with small I/O
sizes & as workloads approached maximum IOPS. These are not the primary benefit derived with Large
write buffer and not shown in detail as a result.
• Note 2: This data should not be used as proxy for drive rebuild time which is based on all the free
resources in a cluster

Internal Use - Confidential 24 Copyright © Dell Inc. All Rights Reserved.

 Creating a vSAN
OSA
configuration
Capacity drives & disk groups

Internal Use - Confidential 25 Copyright © Dell Inc. All Rights Reserved.

 Evaluate benefit of different capacity drives

In All Flash models, the number of SSDs will have an effect on total read throughput and will present
some differences when working with very large IO sizes

Different capacity drives type (SAS,SATA,NVMe, etc) will provide different resync performance

For write intensive workload, a faster capacity tier won’t slow down the rate at which data is moved
from the cache tier to the capacity tier (destaging).

Internal Use - Confidential 26 Copyright © Dell Inc. All Rights Reserved.

 Evaluate benefits of more disk groups per node
The IO capabilities of nodes are directly related to:
• Hybrid vs All Flash
• Number and type of disks in a disk group
• Number of disk groups in a node

If configuration size is being driven by IO requirements,

increasing the number of disk groups per node may
reduce the number of nodes needed

More disk groups will benefit both the Hybrid and All
Flash

Impact of increasing number of disk groups:

• Better performing solution
o Provides performance estimates
• Incremental cost increase

Internal Use - Confidential 27 Copyright © Dell Inc. All Rights Reserved.

 Benefits of more diskgroups per node
Type of workloads that potentially benefit the most from having more disk groups per node:

Video streaming / video Backup solutions where Big Data (Hadoop, Data Warehouse (Oracle
surveillance VxRail is the target Splunk, etc) Data Warehouse, Tableau,
(RecoverPoint, SAP, Microstrategy, etc)
PowerProtect, etc)

More disk groups per node, increases read and write throughput

Performance gain is more noticeable with small block sizes vs large block sizes

Biggest performance achieved when scaling from 1 to 2 diskgroups

Above 3 diskgroups the performance gain is marginal Two disk groups almost doubles the performance for the
incremental cost of a cache drive

Internal Use - Confidential 28 Copyright © Dell Inc. All Rights Reserved.

 Benefits of more diskgroups per node: Mixed workloads

MB/s vs Number of Disk Groups

8000
7103
7000 6865

6000 5532 5479 5622

4896 4953
5000
4251 4399
4081
MB/s

4000
3163
3000 2825 2662
2539
2338
2123 2176
2000 1637
1450 1517 1506
1218
1000 817 742

0
OLTP4K OLTP8K OLTP16K OLTP32K TPCC RDBMS22K

1 DG 2DG 3 DG 4 DG
4xP580N: 4xIntel Xeon Gold 8260L/ Cache Intel P4610 1.6TB/ Capacity Intel P4510
4TB/ 1500GB RAM/ Network 25GbE/ VxRail 7.0.100/RAID1

Internal Use - Confidential 29 Copyright © Dell Inc. All Rights Reserved.

 Benefits of more diskgroups per node: Mixed workloads

IOPS vs Number of Disk Groups

450000
404920 413262
403248
400000 388375
371111
361588 354026350669
350000
311800 311002
299251
300000
272086

250000 236918239763
219692227309
IOPS

209163
192724 197396
200000 179920
162520
150000 140756 140919
128870

100000

50000

0
OLTP4K OLTP8K OLTP16K OLTP32K TPCC RDBMS22K

1 DG 2DG 3 DG 4 DG

4xP580N: 4xIntel Xeon Gold 8260L/ Cache Intel P4610 1.6TB/ Capacity Intel P4510
4TB/1500GB RAM/ Network 25GbE/ VxRail 7.0.100/RAID1

Internal Use - Confidential 30 Copyright © Dell Inc. All Rights Reserved.

 Disk group scaling
Sustained write workloads

VxRail P670F All Flash, 4 Nodes

7,500SW 64K Max, RAID1, no SE, 1-4 DGs per Node Heavy write workloads will benefit
6822
7,000
from more disk groups
6,500
6576 Biggest performance improvement
6,000
5405 when scaling from 1 to 2 DG’s
Throughput MB/s

5,500

5,000
3 and 4 DGs provide similar levels or
4,500
performance
1DG 2DG 3DG 4DG
4,000

3,500
2807
3,000

2,500
0 60 120 180 240 300 360
Test Intervals - 10 seconds Each

Internal Use - Confidential 31 Copyright © Dell Inc. All Rights Reserved.

 Disk group scaling
Peak Performance

VxRail P670F, 25Gb E, 4 No d es

16,000
64K B, RAI D1, 1-4 DGs p er No d e • Read performance is maxed out with
1DG:
14,000 13591 13512 13431
12985 – Network is the first bottleneck to be hit
12,000
• Write performance scales with the
Max Throughput MB/s

10,000 addition of new disk groups

1 DG 2 DGs 3 DGs 4 DGs – Big increase in write performance going from
8,000
1 to 2 DG
6193
5859
6,000 5331

4,000 3323

2,000
RR RW
I/O Type

Internal Use - Confidential 32 Copyright © Dell Inc. All Rights Reserved.

 Number of Disk groups & Performance Quick Facts
Takeaways

1 DG 2 DGs 3 DGs 4 DGs

Biggest When sizing for Performance gains are Apps that require high
performance performance and marginal above 3 DGs throughput will benefit
increase achieved keeping costs low chose for the majority of the most from multiple
when scaling from 2 DGs workloads DGs configs: ex. video
1 DG to 2 DGs streaming, big data,
data warehouses, etc

Internal Use - Confidential 33 Copyright © Dell Inc. All Rights Reserved.

 Mixing disk guidelines in a node for VxRail
Uniformity (homogeneity) is best, but if you must deviate….

Know that performance will be impacted based on the component with the lowest common denominator

Drive Mixing Guideline Cache drives – Capacity drives – Capacity drives –

different disk group same disk group different disk group
Drive size All NVMe – Permitted Permitted* Permitted
(e.g. 1.94TB and 3.84TB)
All Flash – Permitted
Hybrid – Permitted
Protocol type NVMe can replace SAS Not permitted Permitted
(e.g. SAS and SATA) SAS cannot replace NVMe
Flash endurance Permitted Not permitted Not recommended
(e.g. Mixed Use and Write Intensive)
Rotational speed Not Applicable Permitted Permitted
(e.g. 7.2K & 10K HDD)

Reminder:
Hybrid and all-flash configurations cannot be mixed within a cluster or within a node
Dedupe enabled systems must be identical at the disk groups level

Internal Use - Confidential 34 Copyright © Dell Inc. All Rights Reserved.

 Cache & Capacity
Drive Evaluation for
vSAN OSA
Internal Use - Confidential 35 Copyright © Dell Inc. All Rights Reserved.
 Overall drive performance - summary and takeaways
General workload performance and drive types

ENABLING VSAN DATA EFFICIENCY SERVICES? HEAVY WRITE WORKLOADS?

• All-NVMe is the way to go: • All-NVMe is the way to go:

– More IOPS – More sustained throughput
– Significantly lower latency • All-NVMe workloads:
• If All-NVMe is not possible ($), get at least NVMe – Telemetry apps
for the cache tier. – SAP HANA
– Others

NVME AS A CACHE? SAVING $ WITHOUT COMPROMISING PERFORMANCE?

• Big advantage compared with SAS SSD’s cache for • For generic workloads (70/30) with small blocks (<16KB)
workloads that generate medium and large block sizes the performance difference between different type of disk
groups is very small

Internal Use - Confidential 36 Copyright © Dell Inc. All Rights Reserved.

 Medium block tests – OLTP 32 KB
70% Reads, 30% Writes, 100% Random, RAID 1, vSAN Space Efficiency = None
3.4
3.2 All NVMe
3.0 NVMe Cache/SAS Capacity NVMe cache does better
2.8 All SAS for larger block sizes!
2.6
SAS Cache/vSAS Capacity
2.4
SAS Cache/SATA Capacity
Response Time (ms)

2.2
2.0
1.8
1.6
1.4 Normal I/O operating range: 10 – 70% of Max IOPS

1.2
1.0
0.8
0.6
0.4
0 50,000 100,000 150,000 200,000 250,000 300,000 350,000
IOPS

Internal Use - Confidential 37 Copyright © Dell Inc. All Rights Reserved.

 Medium block tests – OLTP 32 KB
70% Reads, 30% Writes, 100% Random, RAID 5, vSAN Space Efficiency = Compression + Dedupe

4.5
All NVMe
NVMe Cache/SAS Capacity All-NVMe if planning to
4.0
All SAS
enable vSAN Data
Services
3.5
SAS Cache/vSAS Capacity
SAS Cache/SATA Capacity
Response Time (ms)

3.0

2.5

2.0

1.5

1.0

0.5
0 50,000 100,000 150,000 200,000 250,000
IOPS

Internal Use - Confidential 38 Copyright © Dell Inc. All Rights Reserved.

 Small block tests – OLTP 4 KB
70% Reads, 30% Writes, 100% Random, RAID 5, vSAN Space Efficiency = None
1.8
All NVMe
1.6 NVMe Cache/SAS Capacity
All SAS
1.4 SAS Cache/vSAS Capacity
Response Time (ms)

SAS Cache/SATA Capacity

1.2

1.0

0.8
For small blocks using
RAID 5, there is only a
0.6 marginal benefit for All-
NVMe in latency and
IOPS.
0.4
0 50,000 100,000 150,000 200,000 250,000 300,000 350,000 400,000 450,000
IOPS

Internal Use - Confidential 39 Copyright © Dell Inc. All Rights Reserved.

 Resync testing results
Overall resync throughput RAID 5, no compression or dedupe
2,500 SW MB/s vendor specs
2,187 Intel NVMe P4510

Samsung SAS PM1643a

2,000

Model
Seagate vSAS
Resync Throughput (MB/s)

Samsung SATA
1,510 1,502
1,500 0 500 1,000 1,500 2,000 2,500 3,000
SW MB/s

1,000 906 Direct correlation between resync

824 performance and specs of capacity
drive type used.

500

Type of cache drive does not affect

resync performance as much.
0
All NVMe NVMe/SAS All SAS SAS/vSAS SAS/SATA

Internal Use - Confidential 40 Copyright © Dell Inc. All Rights Reserved.

 Resync drive performance - summary and takeaways
vSAN resync performance and drive types

WANT FASTER RESYNCS?

SELECT DRIVES THAT HAVE HIGHER WRITE-THROUGHPUT SPECS FOR

THE CAPACITY TIER.
FASTEST VSAN RESYNC PERFORMANCE ACHIEVED ON ALL-NVME.
SLOWEST VSAN RESYNC PERFORMANCE ACHIEVED ON SAS/SATA.
RESYNC PERFORMANCE IS PRIMARILY AFFECTED BY CAPACITY (NOT
CACHE) DRIVE TYPE BUT ALSO BY THE VSAN STORAGE POLICY AND
VSAN DATA SERVICES (DEDUPLICATION & COMPRESSION /
COMPRESSION ONLY / ENCRYPTION)

Internal Use - Confidential 41 Copyright © Dell Inc. All Rights Reserved.

 Diskgroup type takeaways
Takeaways

NVMe Data
Services

NVMe cache for All NVMe helps to

best overall minimize the
performance performance impact
with dedupe

Internal Use - Confidential 42 Copyright © Dell Inc. All Rights Reserved.

 Networking
considerations
10GbE vs 25GbE for OSA

Internal Use - Confidential 43 Copyright © Dell Inc. All Rights Reserved.

 Evaluate benefit of 25GbE vs 10GbE

Chosing 25GbE over 10GbE has considerable benefits on the performance of VxRail:
• Significant increase of read throughput
• Considerable increase of write throughput
• Faster vSAN resyncs
• Reduced upgrade times (due to faster resyncs)
• Improves RAID5 and RAID6 performance
• Minimal per port cost increase for 25GbE over 10GbE
• Cost saving with dual 25GbE vs quad 10GbE

Internal Use - Confidential 44 Copyright © Dell Inc. All Rights Reserved.

 10GbE vs. 25GbE vSAN performance
HOST THROUGHPUT
vSAN 7.0 (RAID-1) 10Gb vs 25Gb
3,500
3,125 3,125 3,125
3,000
2,688 2,713
2,445
2,500

2,000 1,858
MBPS

1,643
1,500 1,250
1,189 1,221

1,000

500

- -
-
10GbE & 2DGS 25GbE & 2DGS 25GbE & 4DGS iPerf

Nominal link speed vSAN SR 64K vSAN SW 64K * 2 (accnt for replic)

• 25GbE enables up to 80% higher read throughput

• With 2 DGs, throughput increase is dependent on capacity drives throughput
• With 4 DGs, 25Gb makes a substantial difference both to reads and writes
• 25GbE makes a significant difference on drive resynchronization throughput, specifically with RAID-5 and RAID-6
• 10GbE can easily be exhausted by vSAN workloads

Internal Use - Confidential 45 Copyright © Dell Inc. All Rights Reserved.

 Significantly faster drive resync throughput
OLTP 4K 15% Workload, no dedupe

14 RAID-5 10GbE 25GbE 14 RAID-6 10GbE 25GbE

Transfer Rate Transfer Rate
763 1784 650 1099
(mbps) (mbps)
12 12

10 10

Response time (ms)

Response time (ms)

8 8

2.3X faster with 25Gb 1.7X faster with 25Gb

6 at lower latency 6

4 4

2 2

0 0
0 140 280 420 560 700 840 980 1120126014001540168018201960210022402380 0 190 380 570 760 950 114013301520171019002090228024702660285030403230

6*P570F-2*Intel 8168-2*DG per host-Cache Samsung NVMe 800GB-3*Capacity Toshiba SAS 7.68TB_10GbE
6*P570F-2*Intel 8168-2*DG per host-Cache Samsung NVMe 800GB-3*Capacity Toshiba SAS 7.68TB_25GbE

Internal Use - Confidential 46 Copyright © Dell Inc. All Rights Reserved.

 10GbE vs 25GbE Quick Facts
Takeaways

Faster vSAN Increased read Increase write Benefits more

resyncs, up to throughput, up to throughput, up to noticeable on multi
diskgroup
2.3X 2X 1.3X configurations

Internal Use - Confidential 47 Copyright © Dell Inc. All Rights Reserved.

 Networking
considerations
25GbE vs 100GbE for vSAN ESA

Internal Use - Confidential 48 Copyright © Dell Inc. All Rights Reserved.

 VxRail ESA
VxRail platforms tested

VXRAIL 8.0.00 ESA VXRAIL 8.0.00 ESA

• 6 x VxRail P670N • 6 x VxRail P670N

• 2x Intel(R) Xeon(R) Gold • 2x Intel(R) Xeon(R) Gold
6330 CPU @ 2.00GHz 6330 CPU @ 2.00GHz
• 28 Cores • 28 Cores

• 1024 GB RAM • 1024 GB RAM

• 2x25GbE (LACP and no • 2x100GbE for VSAN
LACP) for vSAN
• 11x NVMe drives/host:
• 11x NVMe drives/host: • Intel P5600 3.2TB
• Intel P5600 3.2TB
• VxRail version 8.0.000
• VxRail version 8.0.000

These systems are exactly the same: Note: 100GbE requires 100GbE
We just deployed different networking options cables and switching

Internal Use - Confidential 49 Copyright © Dell Inc. All Rights Reserved.

 OLTP 8K 70/30 RAID 6
Compute is the bottleneck on smaller block sizes

OLTP 8K 70/30 RAID 6

25GbE vs 50GbE vs 100GbE
3.00

2.50

2.00
Respone Time (ms)

1.50

1.00

0.50

0.00
0 100000 200000 300000 400000 500000 600000 700000 800000 900000 1000000

IOPS

25GbE 50GbE 100GbE

Internal Use - Confidential 50 Copyright © Dell Inc. All Rights Reserved.

 RDBMS 60/40 22KB RAID5
25GbE vs 50GbE (LACP) vs 100GbE

RDBMS 60/40 22KB RAID5

25GbE vs 50GbE vs 100GbE
Similar peak IOPS for
3.00
LACP and 100GbE
2.52
2.50
2.00

2.00
latency (ms)

30% more IOPS 100GbE vs 25GbE

1.50 AND
14% lower response time
1.00 1.51

0.50

0.00
0 100000 200000 300000 400000 500000 600000 700000
IOPS

25GbE 50GbE 100GbE

Internal Use - Confidential 51 Copyright © Dell Inc. All Rights Reserved.

 RDMS 60/40 22KB RAID6
25GbE vs 50GbE (LACP) vs 100GbE

RDMS 60/40 22KB RAID6

49% more IOPS 100GbE vs 25GbE
25GbE vs 50GbE vs 100GbE AND
23% lower response time
3.00

2.16 2.40
2.50

2.00
1.73
latency (ms)

1.50

1.00

0.50

0.00
0 50000 100000 150000 200000 250000 300000 350000 400000
IOPS
25GbE 50GbE 100GbE

Internal Use - Confidential 52 Copyright © Dell Inc. All Rights Reserved.

 70/30 32KB RAID6
25GbE vs 50GbE (LACP) vs 100GbE for VxRail with vSAN ESA

70/30 32KB RAID6

25GbE vs 50GbE vs 100GbE
3.00
2.49 Throughput wall stops
at ~46% of processing
2.50 2.59 potential

2.00
Latency (ms)

1.32
1.50
78% more IOPS 100GbE vs 25GbE
1.00 AND
19% lower RT
0.50

0.00
0 100000 200000 300000 400000 500000 600000 700000
IOPS

25GbE 50GbE 100GbE

Internal Use - Confidential 53 Copyright © Dell Inc. All Rights Reserved.

 64KB RAID6
Bonded 25GbE LACP is better than 25GbE; 100GbE is preferrable

64KB RAID6
25GbE vs 50GbE vs 100GbE
60000

50000 +89% +82%

+55%
+53%
40000 37094 36739

30579 30886
30000
MB/s

2X or 19651 20185
20000 +82%
+107% +93%
+94%

10000 6453 7352 7508 7480

3538 3860

0
Rand. Write Rand. Read Seq. Write Seq. Read

25GbE 50GbE 100GbE

Internal Use - Confidential 54 Copyright © Dell Inc. All Rights Reserved.

 512KB RAID6
25GbE vs 50GbE (LACP) vs 100GbE

512KB RAID6
25GbE vs 50GbE vs 100GbE
2.3X or
60000 +2.3X or +71%
+126%
+71%
+139%

50000 48594
46705

40000
34831 35358

30000
MB/s

2X or +2X or 20359 20622

20000 +101% +106%
+89% +92%

10000 7422 7615 7494 7630

3693 3967

0
Rand. Write Rand. Read Seq. Write Seq. Read

25GbE 50GbE 100GbE

Internal Use - Confidential 55 Copyright © Dell Inc. All Rights Reserved.

 Key Takeways 100GbE for vSAN ESA
100GbE is the best option for VxRail 8.0 ESA based environments

100GbE provides superior peak performance to both 25GbE and

LACP 25GbE environments.

Be sure customers are advised on how much performance is not

utilized if they insist on 25GbE or LACP with 25GbE, roughly 2x the
nodes to achieve the same performance

The CPU is the constraint for smaller block sizes.

The costs of 100GbE networking will quickly be offset by not needing

to purchase additional nodes in 25GbE designs.

Internal Use - Confidential 56 Copyright © Dell Inc. All Rights Reserved.

 vSAN over RDMA
Internal Use - Confidential 57 Copyright © Dell Inc. All Rights Reserved.
 vSAN Over RDMA
RDMA Overview

RDMA allows servers in a network to exchange

data in memory bypassing the CPU or OS.

RDMA in general improves throughput and

performance as it saves precious CPU cycles

RDMA also reduces the latency on network and

storage applications
Internal Use - Confidential 58 Copyright © Dell Inc. All Rights Reserved.
 Hardware and test environment

4 x VxRail P580N

Intel® Xeon® Platinum 8260L CPU @ 2.40GHz

1512GB RAM

2 Disk Groups / host

1 Cache-Samsung NVMe 800GB
4 Capacity-Intel P4510 4TB

Mellanox Technologies MT27710 Family [ConnectX-4 Lx]

VxRail – 7.0.200; VSAN 7.0U2

Internal Use - Confidential 59 Copyright © Dell Inc. All Rights Reserved.

 Performance Gains with RDMA
OLTP4K (70R/30W 100% Random) - RAID1
RDMA DISABLED RDMA ENABLED

1.20
1.14
1.12
1.10

1.00

0.90 0.85
0.84
Latency (ms)

0.80
0.75
0.70
0.70 0.66
0.65
0.60 0.61
0.60 0.56 0.56
0.52 0.52
0.50 0.48
0.50 0.49 18% more IOPS
0.46
0.45 0.45 20% lower latency
0.40
0 50000 100000 150000 200000 250000 300000 350000 400000 450000 500000

IOPS

Internal Use - Confidential 60 Copyright © Dell Inc. All Rights Reserved.

 Performance gains with RDMA
4xV xR ai l P 580N / 2D G / V xR ai l 7.0.200 / 4xV xR ai l P 580N / 2D G / V xR ai l 7.0.200 /
R andom R eads / R A ID 1 R andom Wri tes / R A ID 1
16000 Improvements across 3000 2841
all block sizes Biggest 2747 2731 2729
13885 14237 2626
14000 13156 improvements
2500 for small blocks
2265
12000 11624
10360 1963
9876 2000
10000
1544
MB/s

MB/s
8000 7097 1500
1184
6000 5343
1000
3698 715
4000 3209
500
2000

0 0
4096 8192 16384 32768 65536 4096 8192 16384 32768 65536

block size (bytes) block size (bytes)

RDMA DISABLED RDMA ENABLED RDMA DISABLED RDMA ENABLED

Internal Use - Confidential 61 Copyright © Dell Inc. All Rights Reserved.

 Improved CPU utilization with RDMA enabled
C P U U tilization during 100% R W 4K @ MA X IOP S during 1 hour
R D MA E nabled vs D isabled
25

20
% CPU Usage

15

10
Up to 5% CPU utilization savings
when enabling RDMA
5

0
0 80 160 240 320 400 480 560 640 720 800 880 960 1040 1120 1200 1280 1360 1440 1520 1600 1680 1760 1840 1920 2000 2080 2160 2240 2320 2400 2480 2560 2640 2720 2800 2880 2960 3040 3120 3200 3280 3360 3440 3520

Elapsed Time (s)

%CPU Usage RDMA Enabled %CPU Usage RDMA Disabled

Internal Use - Confidential 62 Copyright © Dell Inc. All Rights Reserved.

 vSAN over RDMA Quick Facts
Takeaways

Reduces the Saves up to 5% of Increases the maximum Requires RDMA

latency of all the CPU cycles IOPS compatible NICS
workloads
More CPU
available for the
workloads

Internal Use - Confidential 63 Copyright © Dell Inc. All Rights Reserved.

 Memory
considerations
Internal Use - Confidential 64 Copyright © Dell Inc. All Rights Reserved.
 Evaluating memory size options

Different memory configurations are available

Sizer defaults to balanced, lowest cost configuration, meeting capacity requirement
If configuration size is being driven by memory requirements, the use of a larger memory size may reduce the number of
nodes
Impact of increasing memory size:
o Potential for reducing number of nodes and footprint
o A change in configuration options will indicate whether the change has a material effect on cost

Internal Use - Confidential 65 Copyright © Dell Inc. All Rights Reserved.

 Memory Options
• Sizer suggests DIMM size and number of DIMM slots for configuration
• For best performance populate all memory channels
• In the example above, the option “Min memory per processor” was set to 192 to ensure that all memory channels
were populated

Internal Use - Confidential 66 Copyright © Dell Inc. All Rights Reserved.

 Ne
w!
Memory configuration rules for Intel 4th Gen Xeon
Scalable
for 16G Intel Sapphire Rapids based nodes

• 8 memory channels per processor, up to 2 DIMMs

per channel
• Balanced: 16 or 32 DIMMs of equal size per system
– Maximum speed/performance with 16 DIMMs (1 DPC)
– Maximum memory capacity with 32 DIMMs (2 DPC)
– Capacity from 64GB to 2TB per processor
– RDIMM sizes: 16/32/64/128/256 †GB 4800* MT/s

• Near-balanced: Closely match workload requirements

– Four, six or twelve DIMMs of equal size per processor

• VxRail 16G does not offer/support:

– 8GB DIMMs
– Mixing DIMMs VP-760 Memory Layout

– Intel Optane Pmem

– Fourteen DIMM configurations

†
256GB DIMM available only on all-NVMe models
M per
Internal Use -Channel
Confidential(DPC) per CPU 67 Copyright © Dell Inc. All Rights Reserved.
 Intel memory guidelines for 3rd Gen Xeon Scalable

• Wide variety of memory configurations

o 32GB to 2TB per socket

• Balanced 8 and 16 DIMM provide max performance

o Near-balanced 4 or 8 DIMM lose ~30%
o Weigh performance impact against future memory upgrades

• Sizer defaults to balanced configs

o “Enable unbalanced memory” filter to override this

• Greater then 2TB per socket requires L CPU

• PMem Memory Mode can provide greater memory
densities
DIMM per Channel (DPC) 16 GB 128 GB 258 GB
o P and E Series only per CPU RDIMM LRDIMM LRDIMM
o 33% lower price, but with a small impact to performance
1 DPC 128 GB 1024 GB 2048 GB

2 DPC 256 GB 2048 GB 4096 GB

Internal Use - Confidential 68 Copyright © Dell Inc. All Rights Reserved.

 Memory configuration rules
VxRail nodes based on Intel Ice Lake

• Balanced: 8 or 16 DIMMs of equal size per proc.

– Provides for maximum performance
– Capacity from 64GB to 4TB per processor
– RDIMM sizes: 16GB, 32GB, 64GB 3200 MT/s
– LRDIMM size: 128GB, 256GB* 3200 MT/s

• Near-balanced: Closely match workload

• General info:
requirements
– 8 memory channels, with up to two DIMMs per
– Four DIMMs of equal size per processor
channel
– OR – RDIMMs and LRDIMMs must not be mixed
– Mixed DIMM capacities, must populate all DIMM – PMem can mix with either RDIMMs or LRDIMMs
slots
– Bandwidth: 2667 to 3200 MT/s depending on CPU
• For large memory configs PMem in memory mode
• Not supported / available
can be a cost-effective option, without impacting
performance – 8GB RDIMM
– 6 or 12 DIMMs per processor

Internal Use - Confidential 69 Copyright © Dell Inc. All Rights Reserved.

 Ne
Memory configuration rules w!

16G AMD EYPC 4th Gen based nodes

Twelve DDR5 memory channels per processor

• One DIMM per channel (1DPC) supported only

Supported DIMM configurations:

• VE-6615 (Single AMD EPYC):
4, 6, 8, 10 or 12 DIMMs
• VP-7625 (Single AMD EPYC):
4, 6, 8, 10 or 12 DIMMs Single AMD processor - 1 DIMM per memory channel
• VP-7625 (Dual AMD EPYC):
8, 12, 16, 20 or 24 DIMMs
Balanced: For best performance, slotting 12 DIMMs of equal
Supported DIMMs:
size (1DPC) per processor will reap highest memory
• 16GB, 32GB, 64GB, 128GB, 256GB RDIMMs bandwidth of 4800 MT/s
o 256GB DIMM not supported with VP-7625 due
to thermal limitations Near-balanced:
• Capacity from 64GB up to 3TB • Closely match workload requirements
• 4, 6, 8 or 10 DIMMs of equal size per processor
• Mixing DIMM sizes is not supported

Internal Use - Confidential 70 Copyright © Dell Inc. All Rights Reserved.

 P675/E665 memory configuration guidelines
For best performance: For maximum capacity:

Single DIMM per memory channel Two DIMMs per memory channel

• Eight DDR4 memory channels per socket • 4, 8, or 16 RDIMMs of 16GB, 32GB or 64GB in size
• One DIMM per channel provides 3200 MT/s of bandwidth • 128GB LRDIMM available on P675F/N
• Two DIMMs per channel reduces bandwidth to 2933 MT/s • P675F/N supports up to 2TB of system memory
• E665/F supports up to 1TB of system memory
• Not supported: • E665N supports a max of 512GB system memory
• Mixed DIMM sizes
• Populating only 1 or 2 DIMM

Internal Use - Confidential 71 Copyright © Dell Inc. All Rights Reserved.

 AMD Gen 3 memory configuration guidelines
For maximum capacity: 1024GB For best performance: 3200 MT/s
H A H A
DIMM 1 DIMM 0 DIMM 0 DIMM 1 DIMM 1 DIMM 0 DIMM 0 DIMM 1

G B G B
DIMM 1 DIMM 0 DIMM 0 DIMM 1 DIMM 1 DIMM 0 DIMM 0 DIMM 1

F C F C
DIMM 1 DIMM 0 DIMM 0 DIMM 1 DIMM 1 DIMM 0 DIMM 0 DIMM 1

E D E D
DIMM 1 DIMM 0 DIMM 0 DIMM 1 DIMM 1 DIMM 0 DIMM 0 DIMM 1

8 DIMMs in 1 DIMM per channel config 16 DIMMs in 2 DIMM per channel config

AMD supports eight DDR4 memory channels per socket P675 supports 4, 8, or 16 RDIMMs of 16GB, 32GB or 64GB,
• One DIMM per channel provides 3200 MT/s of bandwidth and LRDIMMs of 128GB in size
• Two DIMMs per channel reduces bandwidth to 2933 MT/s • P675 supports up to 2TB of system memory
• 128GB LRDIMM run at 2666 MT/s

E665 supports 4, 8, or 16 RDIMMs of 16GB, 32GB or 64GB in

size
• E665/F supports up to 1TB of system memory
• E665N is limited to up to 512GB of system memory

Internal Use - Confidential 72 Copyright © Dell Inc. All Rights Reserved.

 AMD memory configuration guidelines
Supported, but only recommended if 32 cores or fewer

• Limits memory performance

• Performance Optimized processors
o Optimized for 4 memory channels
o 7232P, 8 core, 3.1GHz, 120W
H A
o 7272, 12 core, 2.9GHz, 120W DIMM 1 DIMM 0 DIMM 0 DIMM 1
o 7282, 16 core, 2.8GHz, 120W
G B
DIMM 1 DIMM 0 DIMM 0 DIMM 1

F C
DIMM 1 DIMM 0 DIMM 0 DIMM 1

E D
DIMM 1 DIMM 0 DIMM 0 DIMM 1

4 DIMMs in 1 DIMM per channel config

Internal Use - Confidential 73 Copyright © Dell Inc. All Rights Reserved.

 General processor
considerations
Internal Use - Confidential 74 Copyright © Dell Inc. All Rights Reserved.
 Optionally: Customize processor model

• If configuration is being driven by CPU requirements, the use of a more powerful CPU may reduce the
node count
• 3rd party software licenses can be based on CPU sockets or processor cores
o Selecting a processor that done not align with software licensing may significantly increase those costs
• Impact of increasing processor capacity:
o Potential for reducing number of nodes and licensing costs
o A change in configuration options will indicate whether the change has a material effect on cost

Internal Use - Confidential 75 Copyright © Dell Inc. All Rights Reserved.

 Choosing between AMD and Intel processors

VxRail offers the option of AMD EPYC2 700x CPU’s in the E and P series:
• Single socket CPU only
• Up to 64 cores per socket (higher than what Intel CPU’s can actually provide)

Consideration needs to be taken Benefits of AMD CPUs: Benefits of Intel CPUs:

when choosing between AMD vs • Higher CPU core count enabling • Larger memory configuration
Intel CPU’s: higher VM consolidation ratios • PMem Support
• Cannot mix AMD and Intel
processors in the same cluster
due to a limitation of the ESXi
hypervisor

Internal Use - Confidential 76 Copyright © Dell Inc. All Rights Reserved.

 Choosing between AMD and Intel processors
Similar performance levels

RDBMS (60/40 22KB)

3.00
3.50

3.00 • Very similar overall performance 2.94

• Intel provides a bit lower latency
2.50 • Both configs provide predictable
performance
Latency (ms)

2.00

1.50

1.00

0.50

0.00
0 20000 40000 60000 80000 100000 120000 140000 160000 180000 200000
IOPS
4 x VxRail E665F 4 x VxRail P670F
2 x Disk Groups per host 2 x Disk Groups per host
1 x AMD EPYC 7763 64-Core CPU @2.45Ghz 2 x Intel(R) Xeon(R) Platinum 8358 32-Core CPU @ 2.60GHz
VMware vSAN 7.0u2 VxRail E665F Milan 64C VxRail P670F Ice Lake 64C VMware vSAN 7.0u2

Internal Use - Confidential 77 Copyright © Dell Inc. All Rights Reserved.

 AMD CPU’s Quick Facts
Takeaways

CPU

High CPU core Main architecture vSAN Performance is

density, up to 64 difference between very similar compared
cores in one single small & large core with latest Intel
socket – achieve configurations is just the generation
higher VM number of CCD’s (Core
consolidation ratios Complex Dies)
per host

Internal Use - Confidential 78 Copyright © Dell Inc. All Rights Reserved.

 VMware Subscription
Licensing considerations
Internal Use - Confidential 79 Copyright © Dell Inc. All Rights Reserved.
 Up
da
VxRail with VMware subscriptions minimum core licensing considerations te
d!

Key VMware subscription licensing updates

Key Broadcom updates effective April 10, 2025 Key updates to Dell quoting in response to Broadcom updates:
 Broadcom has notified the Broadcom VAO partners that starting April 10, 2025, there will  VxRail with Dell-sold VAO subscription quotes with 72 cores or greater
be a minimum purchase quantity of 72 cores per quote for new subscriptions and renewal
transactions. o Dell sellers can continue to transact VxRail quotes with Dell-sold VAO subscriptions that
have at least 72 cores across all the nodes in the quote.
 The 72-core minimum per quote applies to VCF, VVF, vSphere Enterprise Plus, vSphere
Standard, Private AI Foundation, vDefend Firewall, and vDefend Firewall with ATP. Note:
 VxRail with Dell-sold VAO subscription quotes below 72 cores
Currently, VxRail can only be sold with VCF, VVF and vDefend Firewall. VMware products
sold per Virtual Machine or TiB are not subject to the 72-core minimum. o Dell cannot transact VxRail point of sale and renewals quotes with Dell-sold VAO
subscription with less than a total of 72 cores across all nodes in the quote at this time. If a
 The Broadcom minimum requirements do not apply to capacity transactions, where a quote has less than 72 cores it will be rejected and will need to be requoted as a BYOS
customer is buying additional cores for a product that already has an active subscription. Exception RTM transaction or quoted as Dell-sold VAO with additional nodes/cores need
Subscription nodes purchased for a subscription only cluster expansion is a capacity to be added to the quote to meet the 72-core minimum per quote requirement.
transaction.
o In the event a Dell-sold VAO VVF customer does not want to purchase 72 cores for their
 Broadcom subscription renewals will also be subject to the 72-core minimum. VxRail renewal, their only option is to quote BYOS with VCF for all nodes in the cluster.

 Please keep in mind that the 16-core per CPU minimum for VCF and VVF remains in • Please note: Even though Broadcom allows quotes less than 72 cores for subscription only
place. Customers must purchase licenses for the total number of physical cores in each cluster expansions, Dell currently does not support subscription only cluster expansions.
processor, with a minimum of 16 cores per CPU. Communication will be sent when additional information is available.

• Important Notes:
 This policy does NOT apply to the EMEA region. The minimum purchase of Key updates to VxRail Sizing in response to Broadcom updates:
cores per quote will be 16 cores for new subscriptions and renewal transactions
for EMEA-ONLY. The 16-core per CPU minimum for VCF and VVF remains in • No changes to sizing because of 72-core per order minimum policy update.
place.
 All quotes with less than 72 cores previously submitted and rejected in EMEA-
• 16-core per CPU minimum sizing for VCF/VVF remains in place and should be used
ONLY can be resubmitted with less than 72 cores per quote. for sizing total licensed cores needed for VxRail hardware configurations

Internal Use - Confidential 80 Copyright © Dell Inc. All Rights Reserved.

 Up
da
VxRail with VMware subscriptions minimum core licensing considerations te
d!

Key VMware subscription licensing updates FAQs

Q. What are the details of Broadcom’s new policy to require a 72-core minimum purchase quantity per transaction?

A. Broadcom informed VAO partners that starting April 10, 2025, there will be a minimum purchase quantity of 72 cores for new and renewal transactions. The minimum applies to new transactions, where a customer is buying
an impacted SKU for the first time. It does not apply to cluster expansions where a customer purchases additional nodes for a cluster that already has an active subscription. Renewals will also be subject to the 72-core per
transaction minimum.

The Broadcom policy for 72-core minimum per transaction applies to VCF, VVF, vSphere Enterprise Plus, vSphere Standard, Private AI Foundation, vDefend Firewall, and vDefend Firewall with ATP subscription SKUs.
VMware subscription SKUs sold per Virtual Machine or TiB are not subject to the 72-core minimum (i.e., vSAN add on, etc.) Please refer to the VxRail Ordering and eLicensing Guide for details on which VAO subscription
SKUs can be sold with VxRail at this time.

The 16-core per CPU minimum policy for VMware subscription products (e.g. VCF, VVF) remains in place. This means that customers are still required to calculate the cores needed using this formula to determine the total
number of cores needing to be purchased for their order. If the total required cores needed, based on these calculations, are greater than 72 cores, there is no impact on the amount of licensing needing to be purchased
based on this new policy.

Important Notes:
 This policy does NOT apply to the EMEA region. The minimum purchase of cores per quote will be 16 cores for new subscriptions and renewal transactions for EMEA-ONLY. The 16-core per CPU minimum
for VCF and VVF remains in place.
 All quotes with less than 72 cores previously submitted and rejected in EMEA-ONLY can be resubmitted with less than 72 cores per quote.

Q. Does the new Broadcom 72-core minimum requirement per transaction policy apply to all available VxRail with VMware subscriptions Routes to Market?

A. Yes. The increase to the 72-core minimum per transaction applies to VxRail with Dell-sold VAO subscriptions RTM, effective as of April 1, 2025, and VxRail with BYOS Exception RTM, effective as of April 10, 2025.

Q. What do I say if my customer or partner has questions about why Dell is not accepting VxRail with Dell-sold VAO subscription quotes with less than 72 cores?

A. Broadcom is making a change to the minimum number of cores required to purchase VMware subscription software. Dell is aligning our VxRail with Dell-sold VAO subscriptions quoting to the minimum core count required
per quote.

Q. Broadcom has announced that the minimum core count increase will go into effect April 10, 2025. Why is Dell moving to the 72-core minimums before Broadcom?

A. Starting April 10, Broadcom will instate a minimum purchase quantity of 72 cores for new and renewal transactions. To prepare for this, all VxRail with Dell-sold VAO subscription quotes now require a minimum of 72 Cores
total (per Quote). Any VxRail quotes with Dell-sold VAO subscriptions with less than 72 cores not booked by April 1, 2025 are no longer valid and must be requoted to avoid issues.

Internal Use - Confidential 81 Copyright © Dell Inc. All Rights Reserved.

 VxRail Sizing Tool Enhancements – VMware Licensing
• Guided Journey for VMware Subscription Licensing now available:
• Greenfield and Expansions (expanding perpetual with subscription)
• VxRail Ordering & Licensing Guide

• New VxRail Sizing Tool enhancements cater for Greenfield and Expansions
• VxRail Sizing Tool – Release 67 & 69
• Added support for Broadcom Licensing:
• New Licenses section to display the quantity of VMware licenses
required.
• Added toggle include/exclude the VMware licenses pricing in the
sizing calculation.
• Updated VVF license 0.25 TiB per core, as per Broadcom changes

•VxRail Sizing Tool determines VVF/VCF & vSAN Add-on licensing quantities required

•Sizing Enablement videos for VMware subscription licensing available at the

VxRail Technical Webinars site.

Internal Use - Confidential 82 Copyright © Dell Inc. All Rights Reserved.

 VxRail Sizing Tool
What enhancements have been introduced for new VMware subscription licensing requirements?

New toggle to include/exclude Required Licensing quantities New vSAN licensing Tooltip
subscription pricing in cluster (when quantity is negative)
sizing

1. New License filter in Configuration section

• VMware vSphere license auto-selected
New License Filter
from Filters
• Choose VCF or VVF (where applicable)
• New tooltip with VMware licensing KB article
2. New On/Off toggle to include VMware subscription pricing
• Influences outcome of lower cost option if ‘On’
New tooltips 3. Quantity of total VVF/VCF and vSAN TiB licensing
• Note: Any unused vSAN capacity for VCF
will be denoted with a negative number, e.g. -50TiB

VxRail Sizing tool will calculate

VMware subscription licensing
quantities for you!

Internal Use - Confidential 83 Copyright © Dell Inc. All Rights Reserved.

 What VMware software subscription options can I choose for
VxRail?
• VxRail clustered and non-clustered nodes*:
• Requires VMware vSphere Foundation (VVF) subscription licensing at a minimum
or VMware Cloud Foundation (VCF) subscription licensing**
• Applies to following use cases:
• Standard HCI with vSAN (OSA & ESA)
• Dynamic node
• 2-node vSAN
• Satellite nodes
Note: Review VxRail Ordering and Licensing Guide for more detail on each use case for ordering options.

• VCF on VxRail*: Requires VMware Cloud Foundation (VCF) subscription licensing*

*Note: Specific licensing options may vary by customer. Please review the VxRail Ordering and Licensing Guide and contact your Dell/ Broadcom
account team to determine what options are available for your organization.
**Purchase of VCF licensing does not require a full VCF deployment, therefore VCF licensing can be purchased for VxRail with vSAN clusters

Internal Use - Confidential 84 Copyright © Dell Inc. All Rights Reserved.

 Up
d at
ed
VxRail cluster licensing options - Greenfield !
VMware vSphere Foundation (VVF) or VMware Cloud Foundation (VCF)

VVF/VCF core licensing quantities

• Determine which VMware subscription licensing (VVF or VCF) will be required
• VVF and VCF are both Subscription licensing
• Based on Per core pricing
• 16 cores per CPU minimum, e.g. if each CPU has 8 cores, you will require 16.
licenses per CPU
• Number of VVF and VCF core licenses required is based on:

No. cores per CPU * No. CPUs per ESXi host * No. ESXi hosts

vSAN TiB license quantities

• VCF and VVF core licensing provides vSAN capacity entitlements
• vSAN capacity difference for VVF and VCF
• VVF vSAN capacity included in subscription bundle is 0.25 TiB/core
• VCF vSAN capacity included in subscription bundle is 1 TiB/core
• See next slide for calculation and aggregation rules
• If the entitlement is insufficient for the raw capacity of the cluster, then vSAN
Add-on subscription TiBs will be required

VxRail Sizing tool will calculate the VMware subscription licensing quantities for you!
More info: Counting Cores for VMware Cloud Foundation and vSphere Foundation and TiBs for vSAN ([Link])
Internal Use - Confidential 85 Copyright © Dell Inc. All Rights Reserved.
 Differences in vSAN Capacity Types?
Greenfield deployments

vSAN Capacity Type Entitlement/Deployment Details License Criteria

VVF vSAN Capacity • Entitles the customer to vSAN for 0.25 TiB of raw capacity for each VVF
0.25 TiB/Core core
• Capacity can be aggregated and used across VVF core licenses
• Additional capacity can be purchased separately

VCF vSAN Capacity • Entitles the customer to vSAN for 1 TiB of raw capacity for each VCF
1 TiB/Core core • Must license all raw
• Capacity can be aggregated and used across VCF core licenses capacity claimed by
• Additional capacity can be purchased separately vSAN on all hosts in
a cluster

vSAN Add-on 1 • Entitles the customer to vSAN for 1 TiB of raw capacity
1 TiB • 1 vSAN Add-on TiB purchased = 1 TiB of raw capacity
• vSAN Add-on capacity available for VVF and VCF

1 Please review the VxRail Ordering and Licensing Guide and contact your Dell/ Broadcom account team to determine the ordering
process you should follow for your use case.

Internal Use - Confidential 86 Copyright © Dell Inc. All Rights Reserved.

 Greenfield VVF Licensing – Example 1
Number of hosts CPUs/Host Cores/CPU Raw capacity Raw Capacity/ vSAN TiBs (from VVF Subscription Capacity
/host (TiBs) cluster (TiBs) core licenses) Required for vSAN (TiB)
3 1 8 7.68 23.04 12 12

VVF core licensing required = 48 vSAN Add-on subscription TiBs = 12

• 16 cores x 1 CPU x 3 ESXi hosts VVF vSAN capacity entitlement = 0.25 TiB x 48 = 12 TiB
Cluster raw capacity* = 7.68 x 3 = 23.04 TiB
Note: Though the number of cores is 8 on each CPU, customers must purchase 48 Cluster raw capacity (23.04) > vSAN capacity entitlement (12 TiB)
because the minimum subscription capacity is 16 cores per CPU and there is 1 CPU • Exceeds vSAN capacity entitlement, therefore vSAN TiB
on each of the 3 ESXi host. subscription is required
VVF vSAN capacity (12 TiB) < cluster raw capacity (23.04)
Therefore, vSAN Add-on subscription TiB capacity is required based on:
23.04 - 12 = 11.04, roundup
vSAN Add-on subscription TiBs = 12

Greenfield VVF Licensing – Example 2

Number of hosts CPUs/Host Cores/CPU Raw capacity Raw Capacity/ cluster vSAN TiBs (from VVF Subscription Capacity
/host (TiBs) (TiBs) core licenses) Required for vSAN (TiB)
3 2 16 3.84 11.52 24 -12

VVF core licensing required = 96 vSAN Add-on subscription TiBs = 0, excess of 12TiBs

• 16 cores x 2 CPU x 3 ESXi hosts VVF vSAN capacity entitlement = 0.25 TiB x 96 = 24 TiB
Cluster raw capacity* = 3.84 TiB x 3 = 11.52 TiB
VVF vSAN capacity entitlement (24) > Cluster raw capacity (11.52)
• Therefore, no vSAN subscription capacity is required
vSAN Add-on subscription TiBs = 0, with excess of 12TiB

*Based on raw capacity available for vSAN provided by cluster capacity drives
Internal Use - Confidential 87 Copyright © Dell Inc. All Rights Reserved.
 Greenfield VCF Licensing – Example 1
Number of CPUs/Host Cores/CPU Raw capacity/host Raw capacity/cluster vSAN TiBs (from VCF Subscription Capacity
hosts (TiBs) (TiBs) core licenses) Required for vSAN (TiB)

3 1 8 3.840 11.520 48 -36

VCF core licensing required = 48 vSAN Add-on subscription TiBs = 0, excess of 36 TiBs

• 16 cores x 1 CPU x 3 ESXi hosts VCF vSAN capacity entitlement = (1 TiB x 48) = 48 TiB
Cluster raw capacity* = 3.840 x 3 = 11.520 TiB
Note: Though the number of cores is 8 on each CPU, customers must purchase 48 VCF vSAN capacity entitlement (48 TiB) > Cluster raw capacity (11.520)
because the minimum subscription capacity is 16 cores per CPU and there is 1 CPU • Therefore, no vSAN subscription capacity is required
on each of the 3 ESXi host. vSAN Add-on subscription TiBs = 0, with excess of 36TiB

Greenfield VCF Licensing – Example 2

Number of CPUs/Host Cores/CPU Raw capacity/host Raw capacity/cluster vSAN TiBs (from VCF Subscription Capacity
hosts (TiBs) (TiBs) core licenses) Required for vSAN (TiB)

3 2 16 49.920 149.760 96 54

VCF core licensing required = 96 vSAN Add-on subscription TiBs = 54

• 16 cores x 2 CPU x 3 ESXi hosts VCF vSAN capacity entitlement = (1 TiB x 96) = 96 TiB
Cluster raw capacity = 49.920 x 3 = 149.760 TiB
VCF vSAN capacity (96 TiB) < cluster raw capacity (149.760)
• Therefore, vSAN Add-on subscription TiB capacity is required
based on: 149.760 - 96 = 53.76, roundup
vSAN Add-on subscription TiBs = 54

*Based on raw capacity available for vSAN provided by cluster capacity drives

Internal Use - Confidential 88 Copyright © Dell Inc. All Rights Reserved.

 Ne
w!
Expansions – VVF and vSAN subscription licensing
Mixing perpetual and subscription licensing for expanded VxRail clusters

VVF core licensing quantities for expansion nodes

• Subscription VVF core licensing required for new expansion nodes
• Based on Per core pricing
• 16 cores per CPU minimum, e.g. if each CPU has 8 cores, you will require 16 licenses per CPU
• Number of VVF core licenses required is based on:

No. cores per CPU * No. CPUs per ESXi host * No. new ESXi hosts

vSAN Add-on TiB license quantities

• vSAN TiB requirements for mixing perpetual and subscription licensing 3 is based on Expansion rule:
VxRail Sizing tool will # of TiB subscription licenses purchased >= # of physical CPUs in entire vSAN cluster
calculate the VMware • Determine the Raw capacity TiBs for the new expansion nodes available for vSAN
subscription licensing
• Determine Number of physical CPUs across entire cluster, including new expansion nodes
quantities for you!
• Each VVF core license entitles customer to vSAN licensing of 0.25 TiB per core
• If the 0.25 TiB per core entitlement is insufficient, then enough vSAN Add-on licenses will need to be
purchased to address the deficit between the VVF entitlement of 0.25 TiB per core and total TiBs required -
see next slide for examples

1 Please review the VxRail Ordering and Licensing Guide and contact your Dell/ Broadcom account team to determine the ordering process you should follow for your use case.
2 Review KB article Upgrade and Downgrade VMware License Keys ([Link]) for more information relating to Broadcom process of upgrading/ downgrading subscription
license keys
 Ne
Expansions with VVF Licensing – Example 1 w!
Number of CPU per Total CPU Number of Number of Total CPU New Number of cores Total Cores Total Raw capacity Total CPUs in
Existing hosts host Existing hosts new hosts CPU per new hosts per new CPU new hosts new hosts (TiBs) vSAN cluster
host

4 1 4 1 1 1 16 16 2 5
VVF core licensing required = 16 vSAN Add-on subscription licenses = 1
• Customer has an existing 4 node vSAN cluster with a total of 4 physical • Total # CPUs across entire vSAN cluster = (4+1) = 5 CPUs
CPUs (1 per node) • Raw capacity provided with new nodes = 2 TiBs
• 1 new node is being added to existing 4 node cluster • vSAN Expansion rule must be met:
• New node has 1 CPU with 16 cores and 2 TiBs of raw capacity # of TiB subscription licenses purchased >= # of CPUs in vSAN cluster
• 16 VVF core licenses required for new node • 16 VVF core licenses will receive 4 vSAN TiBs towards vSAN capacity licensing (0.25TiB*16)
• VVF core license entitlement is insufficient to meet Expansion rule above (4TiBs < 5 CPU)
Therefore, to meet rule, vSAN Add-on subscription licenses = 1

Expansions with VVF Licensing – Example 2

Number of CPU per Total CPU Number of Number of Total CPU New Number of cores Total Cores Total Raw capacity Total CPUs in
Existing hosts host Existing hosts new hosts CPU per new hosts per new CPU new hosts new hosts (TiBs) vSAN cluster
host

3 2 6 2 2 4 20 80 20 10

VVF core licensing required = 80 vSAN Add-on subscription licenses = 0, excess of 10 TiBs
• Customer has an existing 3 node vSAN cluster with a total of 6 physical CPU • Total # physical CPUs in entire vSAN cluster = 10
(2 CPU per node) • Raw capacity TiB provided by new nodes = 20 TiB
• 2 new nodes are being added to the existing 3 node cluster • vSAN Expansion rule must be met:
• Each new node has 2 physical CPUs, 20 cores per CPU (40 total) and 10 # of TiB subscription licenses purchased >= # CPUs in entire vSAN cluster
TiBs of raw capacity • 80 VVF core licenses will receive 20 vSAN TiBs towards vSAN capacity licensing (0.25TiB*80)
• Number of VVF licenses required = 2 CPUs x 20 cores x 2 hosts • VVF core license entitlement is sufficient to license raw capacity of new nodes, and meets
the Expansion rule above, 20 TiB > 10
• Therefore, vSAN Add-on subscription licenses = 0
Expansions – VCF and vSAN subscription licensing
Mixing perpetual and subscription licensing for expanded VxRail clusters

VCF core licensing2 quantities for expansion nodes

• Subscription VCF core licensing required for new expansion nodes

• Based on Per core pricing:
• 16 cores per CPU minimum, e.g. if each CPU has 8 cores, you will require 16. licenses per CPU
• Number of VCF core licenses required is based on:
No. cores per CPU * No. CPUs per ESXi host * No. new ESXi hosts

vSAN Add-on TiB license quantities

• vSAN TiB requirements for mixing perpetual and subscription licensing 3 is based on Expansion rule:
# of TiB subscription licenses purchased >= # of physical CPUs in entire vSAN cluster
• Determine the Raw capacity TiBs for the new expansion nodes available for vSAN
• Determine Number of physical CPUs across entire cluster, including new expansion nodes
• Each VCF core license entitles customer to vSAN licensing of 1 TiB per core, which oftentimes will be
sufficient to license the Raw capacity of the new expansion nodes and meet Expansion rule
VxRail Sizing tool will • If the 1 TiB per core entitlement is insufficient (which will be less common), then enough vSAN Add-on
calculate the VMware licenses will need to be purchased to address the deficit between the VCF entitlement of 1 TiB per core and
subscription licensing
total TiBs required - see next slide for examples
quantities for you!

1 VCF environment refers to the physical cores in the ESXi hosts where the vSphere in VCF subscription is deployed. Purchase of VCF licensing , does not require a full deployment of VMware Cloud Foundation (platform)
2 Please review the VxRail Ordering and Licensing Guide and contact your Dell/ Broadcom account team to determine the ordering process you should follow for your use case.
3 Review KB article Upgrade and Downgrade VMware License Keys ([Link]) for more information relating to Broadcom process of upgrading/ downgrading subscription license keys
Expansions with VCF Licensing – Example 1
Number of CPU per Total CPU Number of Number of Total CPU New Number of cores Total Cores Total Raw capacity Total CPUs
Existing hosts host Existing hosts new hosts CPU/new hosts per CPU new hosts new hosts (TiBs)
host

3 2 6 2 2 4 24 96 40 10

VCF core licensing required = 96 vSAN Add-on subscription licenses = 0, excess of 46 TiBs
• 2 new nodes will be added to existing 3 node vSAN cluster • Total # CPUs across entire vSAN cluster = (6+4) = 10 CPUs
• Existing cluster has 6 total CPUs • Raw capacity provided with new nodes = 40 TiBs
• Each new node requires: 20TiB raw capacity and 2 CPUs with 24 cores • vSAN Expansion rule must be met:
each # of TiB subscription licenses purchased >= # of CPUs in vSAN cluster
• Scenario meets vSAN Expansion rule (40 TiB > 10 CPUs)
• 96 VCF core licenses, will receive 96 vSAN TiBs towards vSAN capacity licensing
• VCF core license entitlement is sufficient for raw capacity of new nodes (96 TiBs > 40 TiBs).
Therefore, no vSAN Add-on subscription licenses need to be purchased

Expansions with VCF Licensing – Example 2

Number of CPU per Total CPU Number of Number of Total CPU new Number of cores Total cores Total Raw capacity Total CPUs
Existing hosts host Existing hosts new hosts CPU/new hosts per CPU new hosts new hosts (TiBs)
host
3 2 6 3 1 3 16 48 60 9

VCF core licensing required = 48 vSAN Add-on subscription licenses = 12

• 3 new nodes will be added to existing 3 node vSAN cluster • Total # CPUs across entire vSAN cluster = (6+3) = 9 CPUs
• Existing cluster has 4 total CPUs • 48 VCF core licenses, will receive 48 SAN TiBs towards vSAN capacity licensing
• Each new node requires: 20TiB raw capacity and 1 CPU with 16 cores • Raw capacity provided with new nodes = 60 TiBs
• vSAN Expansion rule must be met:
# of TiB subscription licenses purchased >= # of CPUs in vSAN cluster
• VCF vSAN capacity entitlements is sufficient to meet vSAN exception rule (48 TiBs > 9 CPUs)
BUT will not accommodate all of the raw capacity of the new nodes (48 TiBs < 60 TiBs).
Therefore, 12 vSAN Add-on subscription licenses must be purchased (60 TiBs – 48 TiBs)
 Up
VxRail & VMware subscription licensing d at
ed
!

Review VxRail Ordering

Use the VxRail Sizing Strike the best balance Generate quote in
& Licensing Guide to
Tool to generate VxRail between cost and Ordering Tools
define the VMware
configuration and performance of your VxRail
licensing type required
VMware subscription configuration
(see Resources slide)
licensing quantities

VMware vSAN Capacity entitlement

Deployment Core licenses
subscription with VVF/VCF Core vSAN Add-on subscription licenses
Type Quantity New Nodes
licensing type licensing
vSAN clusters exceeding a total of 0.25 TiB per VCF core requires a purchase of vSAN Add-
Greenfield VVF 1 core = 0.25 TiB on licenses that will accommodate the excess TiBs required to meet the raw capacity of the
vSAN cluster

vSAN clusters exceeding a total of 1 TiB per VCF core requires a purchase of vSAN Add-on
Greenfield VCF 1 core = 1 TiB licenses that will accommodate the excess TiBs required to meet the raw capacity of the
vSAN cluster
(No. core licenses per
CPU*) x (No. CPUs per # of TiB subscription licenses purchased >= # of CPUs in entire vSAN cluster.
Expansions - ESXi host) x In addition, vSAN clusters exceeding a total of 0.25 TiB per VVF core requires a purchase of
Mixing with VVF (No. ESXi hosts) 1 core = 0.25 TiB
vSAN Add-on licenses that will accommodate the excess TiBs required to meet the raw
perpetual
capacity of the vSAN cluster, if the exception rule is not met.

# of TiB subscription licenses purchased >= # of CPUs in entire vSAN cluster.

Expansions -
In addition, vSAN clusters exceeding a total of 1 TiB per VCF core requires a purchase of
Mixing with VCF 1 core = 1 TiB
*Minimum 16 cores vSAN Add-on licenses that will accommodate the excess TiBs required to meet the raw
perpetual per CPU requirement capacity of the vSAN cluster, if the exception rule is not met.
 Up
d at
VMware subscription licensing resources ed
!

Resource URL

VxRail Ordering & Licensing Guide VxRail Ordering & Licensing Guide

VxRail Technical Webinar Sizing Enablement VxRail Technical Webinars

VMware Pricing and Packaging Overview

VMware Subscription Licensing P&P Overview

VMware Subscription Licensing P&P FAQ VMware Pricing and Packaging FAQ

VxRail VMware Subscription Licensing FAQ FAQs: VxRail with VMware Subscriptions

Upgrade/Downgrade KB Article Broadcom KB Article: Upgrade and Downgrade VMware License Keys

Internal Use - Confidential 94 Copyright © Dell Inc. All Rights Reserved.

 Up
d at
ed

16G VxRail platforms

Performance of 16G VxRail with AMD EPYC 4th Generation
processors

Internal Use - Confidential 95 Copyright © Dell Inc. All Rights Reserved.

 AMD 4th Gen EPYC™ Processors on 16G VxRail
Built with 5nm Technology

Next Generation AMD Processors Memory: DDR5

• AMD 4th Generation EPYC (Genoa) • DDR5 (4800MT/s)
 Single or Dual AMD EPYC on 16G  Latest DRAM technology with higher
speed & Bandwidth
VxRail  Twelve DDR5 4800MHz memory
 SP5 Socket, 5nm technology channels per processor
 Up to 96 “Zen 4” cores per socket  Improved RAS features with on-die ECC

Gen over Gen AMD EPYC comparison

PCIe Gen5 Capability • Dual AMD EPYC (new!) supported with VP-
• Doubles throughput compared to PCIe 7625
Gen4 • Up to 3x more CPU cores
• 50% more memory channels
 Up to 128 PCIe lanes per CPU
• 50% increased memory speed
 Benefits NVMe drives, GPUs, • Up to 3x increase in total system memory
networking cards supporting Gen5 • 38% more IOPS1, 30% smaller cluster
bandwidth footprint2
1
Based on internal Dell testing, March 2024, comparing 15G VxRail (Gen 3 AMD EPYC 7763 CPU, 64c) with 16G VxRail (Gen 4 AMD EPYC 9534 processors, 64c) RAID
5 configuration for OLTP workload.
2
Based on internal Dell analysis, March 2024
Internal Use - Confidential 96 Copyright © Dell Inc. All Rights Reserved.
 Ne
V X R A I L 1 6 T H G E N E R AT I O N A M D P L AT F O R M S w!

VP-7625 with vSAN ESA

vSAN OSA Vs ESA comparison

Internal Use - Confidential 97 Copyright © Dell Inc. All Rights Reserved.

 Ne
w!

VxRail VxRail VP-7625 All-NVMe VxRail

8.0.300 7.0.510
• 6 x VxRail VP-7625 • 6 x VxRail VP-7625
• 2 x AMD EPYC 9354 CPU • 2 x AMD EPYC 9354 CPU
(32C @ 3.25GHz) (32C @ 3.25GHz)
 64 cores per node  64 cores per node

• 1024 GB RAM • 1024 GB RAM

• 2 x Broadcom SFP28 • 2 x Broadcom SFP28
Ethernet OCP 3.0 Ethernet OCP 3.0
Adapter 25 Gbit/s Adapter 25 Gbit/s
• 20 drives per node: • 4 DG per host:
• Dell Ent NVMe PM1733a RI • 1x cache Dell Ent NVMe
3.84TB CM6 MU 1.6TB KIOXIA

• vSAN ESA 8.0 Update 3a • 4x capacity Dell Ent NVMe

PM1733a RI 3.84TB

• vSAN OSA 7.0 update 3

Internal Use - Confidential 98 Copyright © Dell Inc. All Rights Reserved.

98
 Ne
w!

Small I/O
Mixed
Workload
70% random reads
30% random writes
• RAID 6 ESA provides 2.98x
more IOPS and 70% reduction
Space efficiency = none in latency compared to RAID 6
OSA
• Steady curve for ESA latency
with ~0.5ms response time for

RAID 6 @ max 1.3million IOPS

OSA Vs ESA

Internal Use - Confidential 99 Copyright © Dell Inc. All Rights Reserved.

 Ne
w!

Small I/O
Mixed
Workload
70% random reads
• RAID 5 ESA provides 2.18x
30% random writes more IOPS and 53% less
latency compared to RAID 5
Space efficiency = none OSA
• Steady curve for ESA latency
with <0.5ms response time
@max 1.4million IOPS

RAID 5
OSA Vs ESA

Internal Use - Confidential 100 Copyright © Dell Inc. All Rights Reserved.
 Ne
w!
VP-7625, 6 nodes, All-NVMe, OLTP 4k
CPU Usage MHz, OSA Vs ESA, RAID 5

Small I/O 98% more CPU usage with

742298 MHz

Mixed ESA, results in 2.18x more

IOPS than OSA

Workload 373247 MHz

ESA, R5

70% random reads OSA, R5

30% random writes

Space efficiency = None VP-7625, 6 nodes, All-NVMe, OLTP 4k
CPU Usage MHz, OSA Vs ESA, RAID 6

769254 MHz
CPU comparison

107% more CPU usage with

RAID 5/6 ESA, results in 2.18x more

IOPS, compared to OSA 370226 MHz
ESA, R6
OSA Vs ESA OSA, R6

Internal Use - Confidential

As IOPs increases CPU also increases
101 Copyright © Dell Inc. All Rights Reserved.
 Ne
w!

Medium
I/O Mixed
Workload
70% random reads
30% random writes
RAID 6 ESA provides 26% more
Space efficiency = none IOPS and 36% less latency
compared to RAID 6 OSA

RAID 6
OSA Vs ESA

Internal Use - Confidential 102 Copyright © Dell Inc. All Rights Reserved.
 Ne
w!

Medium
I/O Mixed
Workload
70% random reads
RAID 5 ESA provides 16% more
30% random writes IOPS and 16% less latency,
compared to RAID 5 OSA
Space efficiency = none

RAID 5
OSA Vs ESA

Internal Use - Confidential 103 Copyright © Dell Inc. All Rights Reserved.
 Ne
w!
VP-7625, 6 nodes, All-NVMe, OLTP 32k

Medium
CPU Usage MHz, OSA Vs ESA, RAID 5

I/O 63% more CPU usage with

ESA, results in 16% more IOPS

Mixed
than OSA
536000 MHz

Workload ESA, R5
327978 MHz

70% random reads

OSA, R5

30% random writes

Space efficiency = None
VP-7625, 6 nodes, All-NVMe, OLTP 32k
CPU Usage MHz, OSA Vs ESA, RAID 6

CPU comparison 59% more CPU usage with

ESA, results in 26% more IOPS 539872 MHz

than OSA

RAID 5/6 337487 MHz

ESA, R6
OSA Vs ESA OSA, R6

As IOPs increases CPU also increases

Internal Use - Confidential 104 Copyright © Dell Inc. All Rights Reserved.
 Ne
w!

Large I/O
Mixed
Workload RAID 6 ESA provides 11.9%
more IOPS and 8.7% less
latency compared to RAID 6
OSA
70% random reads
30% random writes
Space efficiency = none

RAID 6
OSA Vs ESA

Internal Use - Confidential 105 Copyright © Dell Inc. All Rights Reserved.
 Ne
w!

Large I/O
Mixed
Workload RAID 5 ESA provides 11% more
IOPS and 8.7% reduction in
latency compared to RAID 5 OSA
70% random reads
30% random writes
Space efficiency = none

RAID 5
OSA Vs ESA

Internal Use - Confidential 106 Copyright © Dell Inc. All Rights Reserved.
 Ne
VP-7625, 6 nodes, All-NVMe, OLTP 64k w!
CPU Usage MHz, OSA Vs ESA, RAID 5

Large I/O 81% more CPU usage with

ESA, results in 2.5% more

Mixed IOPS than OSA

408971 MHz

Workload 225557 MHz

ESA, R5

OSA, R5
70% random reads
30% random writes
Space efficiency = None VP-7625, 6 nodes, All-NVMe, OLTP 64k
CPU Usage MHz, OSA Vs ESA, RAID 6

82% more CPU usage with

CPU comparison
ESA, results in 11.19% 419815 MHz

more IOPS than OSA

RAID 5/6 230003 MHz

ESA, R6

OSA Vs ESA OSA, R6

As IOPs increases CPU also increases

Internal Use - Confidential 107 Copyright © Dell Inc. All Rights Reserved.
 Ne
w!

Sustained ESA, R5

Write 64K OSA, R5

34% more CPU achieves 15%
more throughput

CPU comparison
Space efficiency = none

RAID 5/6 ESA, R6

44% more CPU usage
OSA Vs ESA achieves 30% more throughput
OSA, R6

Same pattern increased CPU based on increased throughput

Internal Use - Confidential 108 Copyright © Dell Inc. All Rights Reserved.
 Ne
Peak performance - Small and Large I/O size w!

Maximum IOPS, space efficiency = none

4K I/O, IOPS 64K I/O, throughput

+18% with R6 ESA

2.9x increase
with R5 ESA
+24% with R6 ESA

3.97x increase
2.5x increase with R6 ESA
with R5 ESA
3.89x increase
with R6 ESA

For small I/O, ESA is significantly more performant than OSA, with up to For large I/O, similar performance between ESA and OSA, with OSA
3.97x increase in IOPS with RAID 6 for Sequential Writes, and 2.9x providing marginal improved throughput (~1-2%), with exception of
increase in IOPS with RAID 5 for Sequential Writes. Sequential Writes where ESA is most performant with 13% more throughput

Internal Use - Confidential 109 Copyright © Dell Inc. All Rights Reserved.
 Ne
w!

Key • As expected, ESA outperforms OSA for small, medium and large I/O mixed
workloads
Takeaways • Biggest gains seen with small I/O, with RAID 6 ESA providing 2.98x
for 16th more IOPS and 70% reduction in latency compared to RAID 6 OSA
• Steady increase in latency with <0.5ms @ max 1.4 million IOPS
Generation
• While CPU usage is driven higher as IOPS increase, greater throughput is
AMD VxRail achieved across all I/O block sizes with ESA
with vSAN • 99% more CPU usage for small I/Os, results in 2.18x more IOPS

ESA • For peak performance tests, ESA performed significantly better than OSA for
small I/O block size, in particular for Random and Sequential writes
• 3.97x higher IOPS with R5 ESA for sequential writes
• Very similar performance between OSA and ESA for large I/O block
size, with exception of sequential writes workload where ESA
outperformed OSA by 13% for sequential writes

• ESA provided more throughput for sustained sequential writes with large I/O:
• 30% more throughput provided by RAID 6 ESA, compared to OSA

Internal Use - Confidential 110 Copyright © Dell Inc. All Rights Reserved.
 V X R A I L 1 6 T H G E N E R AT I O N A M D P L AT F O R M S

VE-6615
Storage policy
comparison
Internal Use - Confidential 111 Copyright © Dell Inc. All Rights Reserved.
 VxRail VxRail VE-6615 All-Flash
7.0.510

• 6 x VxRail VE-6615
• 1 x AMD EPYC 9534,
2.45GHz
• 64 cores per node
• 512 GB RAM
• 2 x Broadcom SFP28
Ethernet OCP 3.0 Adapter
25 Gbit/s
• 2 DG per host:
• 1 x Cache, SAMSUNG
MZILG1T6HCJRAD3 1.6 TB
• 4 x Capacity, KIOXIA
KPM7XRUG3T84 3.8 TB

• vSAN OSA 7.0 U3

Internal Use - Confidential 112 Copyright © Dell Inc. All Rights Reserved.
112
 Small I/O
Mixed
Workload
70% random reads
30% random writes
Space efficiency = none

Internal Use - Confidential 113 Copyright © Dell Inc. All Rights Reserved.
 Medium
I/O Mixed
Workload
70% random reads
30% random writes
Space efficiency = none

Internal Use - Confidential 114 Copyright © Dell Inc. All Rights Reserved.
 Large I/O
Mixed
Workload
70% random reads
30% random writes
Space efficiency = none

Internal Use - Confidential 115 Copyright © Dell Inc. All Rights Reserved.
 Large I/O
max Destaging

sustained
sequential
writes • Similar performance
between RAID 1 and
RAID 5

AKA “steady state” • Average of 43% more

throughput with RAID 1
tests, and RAID 5 compared to
RAID 6

space efficiency =
no SE

Throughput MB/s

Internal Use - Confidential 116 Copyright © Dell Inc. All Rights Reserved.
 Peak performance - Small and Large I/O size
Maximum IOPS, space efficiency = none

For both I/O Sizes:

Similar performance
for Random Read for
all storage policies,
with expected drop
from RAID 1 to RAID
6 for Random Write

4K I/O 64K I/O

Internal Use - Confidential 117 Copyright © Dell Inc. All Rights Reserved.
 V X R A I L 1 6 T H G E N E R AT I O N A M D P L AT F O R M S

VP-7625
Storage policy
comparison
Internal Use - Confidential 118 Copyright © Dell Inc. All Rights Reserved.
VxRail VxRail VP-7625 All-NVMe
7.0.510
• 6 x VxRail VP-7625
• 2 x AMD EPYC 9354 CPU
(32C @ 3.25GHz)
 64 cores per node

• 1024 GB RAM
• 2 x Broadcom SFP28
Ethernet OCP 3.0
Adapter 25 Gbit/s
• 4 DG per host:
• 1x cache Dell Ent NVMe
CM6 MU 1.6TB KIOXIA
• 4x capacity Dell Ent NVMe
PM1733a RI 3.84TB

• vSAN OSA 7.0 Update 3

Internal Use - Confidential 119 Copyright © Dell Inc. All Rights Reserved.
119
 Small I/O
Mixed
Workload
70% random reads
30% random writes
Space efficiency = none

Internal Use - Confidential 120 Copyright © Dell Inc. All Rights Reserved.
 Medium
I/O Mixed
Workload
70% random reads
30% random writes
Space efficiency = none

Internal Use - Confidential 121 Copyright © Dell Inc. All Rights Reserved.
 Large I/O
Mixed
Workload
70% random reads
30% random writes
Space efficiency = none

Internal Use - Confidential 122 Copyright © Dell Inc. All Rights Reserved.
 Large I/O
max RAID 1 destaging

sustained
sequential • Up to RAID 1 destaging, RAID 1

writes provides average of 23% more

throughput than RAID 5
• RAID 5 provides up average of 37%
more throughput than RAID 6
AKA “steady state” • RAID 5/6 destaging not shown in this
tests, test, expected post-60 minutes

space efficiency =
no SE

Throughput MB/s

Internal Use - Confidential 123 Copyright © Dell Inc. All Rights Reserved.
 Peak performance - Small and Large I/O size
Maximum IOPS, space efficiency = none
For both I/O Size:
Similar performance
for Random Read for
all storage policies,
with expected drop
from RAID 1 to RAID
6 for Random Write

4K I/O 64K I/O

Internal Use - Confidential 124 Copyright © Dell Inc. All Rights Reserved.
Key
Takeaways • Similar performance differences of Raid 1, 5, and 6 are achieved
compared to other VxRail vSAN OSA clusters for VE-6615 & VP-7625
for 16th • VE-6615: Up to 789,000 IOPS with RAID 1 and sub millisecond
Generation response time for small block I/O
VxRail with • For sustained sequential writes with large I/O:
AMD: • The VE-6615 delivered similar performance up to destaging point
for RAID 1 and RAID 5
VE-6615 & • The VP-7625 delivered more throughput performance for RAID 1
VP-7625 compared to RAID 5 and RAID 6

• Similar performance for Random Reads for all storage policies with
expected drop from RAID 1 to RAID 6 for Random Writes

Internal Use - Confidential 125 Copyright © Dell Inc. All Rights Reserved.
 VxRail AMD EPYC Generation
Comparison
VE-6615 Vs E665F

Internal Use - Confidential 126 Copyright © Dell Inc. All Rights Reserved.
 VxRail AMD Platforms tested
Generation over Generation VxRail AMD EPYC

16th Generation 15th Generation

• 4 x VxRail VE-6615 • 4 x VxRail E665F
• 1 x AMD EPYC 9534 • 1 x AMD EPYC 7763
CPU @ 2.45 GHz CPU @ 2.45GHz
• 64 cores per node  64 cores per node

• 512 GB RAM • 512 GB RAM

• 2 x Broadcom SFP28 • 2 x Broadcom BCM57414
Ethernet OCP 3.0 Adapter Ethernet 25 Gbit/s
25 Gbit/s
• 2 DG per host:
• 2 DG per host: • 1 x Cache, SAMSUNG
• 1 x Cache, SAMSUNG MZILT1T6HBJR0D3 1.6
MZILG1T6HCJRAD3 1.6 TB TB
• 4 x Capacity, KIOXIA • 4 x Capacity, Western
KPM7XRUG3T84 3.8 TB Digital 3.84 TB
WUSTR1538ASS200
• vSAN OSA 7.0 Update 3o
• vSAN OSA 7.0 Update 2

Internal Use - Confidential 127 Copyright © Dell Inc. All Rights Reserved.
 Small I/O
mixed
workload
28% less latency @
70% reads 75% IOPS

30% writes
+20% more IOPS
No space efficiency 28% Lower latency

RAID 1 comparison

Internal Use - Confidential 128 Copyright © Dell Inc. All Rights Reserved.
 Small I/O
mixed
workload
38% less latency
@ 75% IOPS

70% reads
30% writes
+21% more IOPS
No space efficiency 38% lower latency

RAID 5 comparison

Internal Use - Confidential 129 Copyright © Dell Inc. All Rights Reserved.
 Medium I/O
mixed
workload

70% reads
30% writes 26% less latency
@ 75% IOPS
No space efficiency +8% more IOPS
26% lower latency
RAID 1 comparison

Internal Use - Confidential 130 Copyright © Dell Inc. All Rights Reserved.
 Medium I/O
mixed
workload

70% reads
30% writes
No space efficiency
RAID 5 comparison

Internal Use - Confidential 131 Copyright © Dell Inc. All Rights Reserved.
 Large I/O
mixed
workload

70% reads
30% writes
No space efficiency
RAID 5 comparison

Internal Use - Confidential 132 Copyright © Dell Inc. All Rights Reserved.
 Peak performance - Small and Large I/O size, RAID 1
Maximum IOPS, space efficiency = none, VxRail VE-6615 Vs E665F

8K I/O RAID 1 64K I/O RAID 1

Internal Use - Confidential 133 Copyright © Dell Inc. All Rights Reserved.
 Peak performance - Small and Large I/O size, RAID 5
Maximum IOPS, space efficiency = none, VxRail VE-6615 Vs E665F

8K I/O RAID 5 64K I/O RAID 5

Internal Use - Confidential 134 Copyright © Dell Inc. All Rights Reserved.
 Large I/O
max
sustained
sequential
writes

RAID 5
AKA “steady state”
tests,
Space efficiency =
none

Throughput MB/s
Internal Use - Confidential 135 Copyright © Dell Inc. All Rights Reserved.
 Key • For small and medium I/O block size, the new VxRail VE-6615 (Gen 4
Takeaways AMD EPYC) outperforms VxRail E665F significantly for both max
IOPS/throughput and latency
for 16th
o VE-6615 provides up to 20% more IOPS, 28% lower latency for
Generation small I/O mixed workload, RAID 1
VxRail with o VE-6615 provides up to 21% more IOPS, 38% lower latency for
AMD medium I/O mixed workload, RAID 5
o VE-6615 provides up to 17% more throughput for random reads

• At larger block size the performance is more equal in terms of IOPS, with
the VE-6615 achieving significantly lower latency than the E665F
o Up to 6.5% more IOPS with an average of 40% lower latency, for
large I/O mixed workload, RAID 5

Internal Use - Confidential 136 Copyright © Dell Inc. All Rights Reserved.
 Ne
w!

VD-4000 with vSAN 8.0

vSAN OSA and ESA

Internal Use - Confidential 137 Copyright © Dell Inc. All Rights Reserved.
 V X R A I L 1 5 T H G E N E R AT I O N P E R F O R M A N C E

VxRail VD-4000 with vSAN OSA

Internal Use - Confidential 138 Copyright © Dell Inc. All Rights Reserved.
 VxRail VD-4000 using VD-4520c nodes
8.0.210

• 6 x VxRail VD-4520c
• 1 x Intel® Xeon® D-2796NT
CPU @ 2.0 GHz
 20 Cores

• 512 GB RAM
• Intel® E823-C 25 GbE
• 2 DGs per host:
 1 x Cache, SK Hynix PE8030
MU NVMe - 800 GB
 3 x Capacity, SK Hynix PE8110
RI NVMe – 3.84 TB

• vSAN 8.0 U2

Internal Use - Confidential 139 Copyright © Dell Inc. All Rights Reserved.
 Small I/O
mixed
workload
70% random reads
30% random writes
Space efficiency:
• As expected with vSAN
Compression and OSA, RAID 1 performed
deduplication best
• RAID 5 averaged 19%
lower IOPS and 26%
higher latency compared
to RAID 1

Internal Use - Confidential 140 Copyright © Dell Inc. All Rights Reserved.
 Medium
I/O mixed
workload
70% random reads
30% random writes
Space efficiency:
Compression and
deduplication

• As expected, RAID 1
performed best
• RAID 5 averaged 17%
lower IOPS and 28%
higher latency
compared to RAID 1

Internal Use - Confidential 141 Copyright © Dell Inc. All Rights Reserved.
 CPU Usage

Medium
I/O mixed
workload
• RAID 1 drove highest CPU usage
• RAID 5 averaged 11% lower CPU
usage compared to RAID 1
• RAID 6 averaged 16% lower CPU
usage compared to RAID 1
70% random reads
30% random writes
Space efficiency:
Compression and
deduplication

Internal Use - Confidential 142 Copyright © Dell Inc. All Rights Reserved.
 Large I/O
mixed
workload
70% random reads
30% random writes
Space efficiency:
Compression and
deduplication
• As expected, RAID 1
performed best
• RAID 5 averaged 17%
lower IOPS and 28%
higher latency compared
to RAID 1

Internal Use - Confidential 143 Copyright © Dell Inc. All Rights Reserved.
Key • For the VD-4520c, similar performance differences are seen with
Takeaways RAID 1, 5 and 6 that are seen with other VxRail clusters with vSAN
OSA with RAID 1 being be the most performant
for VD-4000
• RAID 5 provides 26% lower IOPS and 30% higher latency
with VxRail compared to RAID 1 for medium I/O block size mixed workload
8.0.210: • For small I/O block size, RAID 5 averaged 19% lower IOPS and
26% higher latency compared to RAID 1

vSAN OSA • RAID 1 drives CPU usage higher compared to other storage policies
for medium I/O block size
• RAID 5 averaged 11% lower CPU usage than RAID 1
• RAID 6 averaged 16% lower CPU usage than RAID 1

Internal Use - Confidential 144 Copyright © Dell Inc. All Rights Reserved.
 V X R A I L 1 5 T H G E N E R AT I O N P E R F O R M A N C E

VxRail VD-4000 with vSAN ESA

Internal Use - Confidential 145 Copyright © Dell Inc. All Rights Reserved.
 VxRail VD-4000 using VD-4520c nodes
8.0.210

• 6 x VxRail VD-4520c
• 1 x Intel® Xeon® D-2796NT
CPU @ 2.0 GHz
 20 Cores

• 512 GB RAM
• Intel® E823-C 25 GbE
• 8 x SK Hynix PE8110 RI
NVMe capacity drives per
node – 3.84 TB
• vSAN ESA 8.0 U2

Internal Use - Confidential 146 Copyright © Dell Inc. All Rights Reserved.
 Micron 7400 only ESA drives supported for 8.0.210
• Testing completed with SK Hynix drives before support for these drives were dropped
• Performance with Micron drives would be expected to be higher than results achieved with the SK Hynix drives, as per
Manufacturer charts

Internal Use - Confidential 147 Copyright © Dell Inc. All Rights Reserved.
 Small I/O
mixed
workload
70% random reads
30% random writes
Space efficiency: As expected with vSAN ESA,
RAID 5 performs best
• RAID 5 achieved an
Compression
average of 10% more IOPS
and 5% lower latency
compared to RAID 1

Internal Use - Confidential 148 Copyright © Dell Inc. All Rights Reserved.
 Medium
I/O mixed
workload
70% random reads
30% random writes
As expected with vSAN ESA,
RAID 5 performs best:
Space efficiency: • RAID 5 averaged 11%
more IOPS and 8% lower
Compression latency compared to
RAID 1
• RAID 6 and RAID 1
similar performance

Internal Use - Confidential 149 Copyright © Dell Inc. All Rights Reserved.
 CPU Usage

Medium Similar CPU usage for all three storage

policies RAID 1 drove lowest CPU

I/O mixed usage

• RAID 5 averaged 2% higher CPU

workload usage compared than RAID 1

• RAID-6 averaged 3% higher CPU
usage vs RAID 1

70% random reads

30% random writes
Space efficiency:
Compression

Internal Use - Confidential 150 Copyright © Dell Inc. All Rights Reserved.
 Large I/O
mixed
workload
70% random reads
30% random writes
Space efficiency:
Compression
As expected with vSAN
ESA, RAID 5 performs best
• RAID 5 averaged 6%
more IOPS and 9%
lower latency compared
to RAID 1
• RAID 6 and RAID 1
perform similarly

Internal Use - Confidential 151 Copyright © Dell Inc. All Rights Reserved.
Key • As expected with vSAN ESA, RAID 5 performs best
Takeaways
• On average, RAID 5 provides 11% more IOPS and 8% lower
for VD-4000 latency compared to RAID 1 for medium I/O block size mixed
with VxRail workload

8.0.210: • RAID 6 and RAID 1 have similar performance with RAID 6 being
the least performant of the two

vSAN ESA • On average, RAID 6 provides 5% higher latency and 4% less

IOPS compared to RAID 1 for small I/O block size mixed
workload

• Similar CPU usage for all three storage policies with RAID 1
driving lowest CPU usage

Internal Use - Confidential 152 Copyright © Dell Inc. All Rights Reserved.
 V X R A I L 1 5 T H G E N E R AT I O N P E R F O R M A N C E

VxRail VD-4000 platforms:

vSAN OSA Vs ESA
Internal Use - Confidential 153 Copyright © Dell Inc. All Rights Reserved.
 Small I/O
mixed • ESA achieves 35% more
IOPS than OSA

workload
• ESA delivers 31% lower
latency than OSA (@ OSA
max IOPS)

RAID 5
70% reads / 30% writes
Space efficiency:
• vSAN OSA –
Compression & 31 % lower latency at OSA
deduplication max IOPS

• vSAN ESA –
Compression only

Internal Use - Confidential 154 Copyright © Dell Inc. All Rights Reserved.
 Medium vSAN ESA performs better

I/O mixed than vSAN OSA

• Average 2% more IOPS

workload • Average 22% lower

latency

RAID 5
70% reads / 30% writes
Space efficiency:
• vSAN OSA –
Compression &
deduplication
• vSAN ESA –
Compression only

Internal Use - Confidential 155 Copyright © Dell Inc. All Rights Reserved.
 Large I/O • Overall, OSA performs better than

mixed ESA
• 11% more IOPS and average

workload •
12% lower latency
vSAN ESA delivers an average of
21% lower latency @10% - 50%
IOPS
• vSAN ESA drives 57% higher

RAID 5 latency @70% - 100% IOPS

70% reads / 30% writes

Space efficiency:
• vSAN OSA –
Compression &
deduplication
• vSAN ESA –
Compression only

Internal Use - Confidential 156 Copyright © Dell Inc. All Rights Reserved.
 Large I/O
mixed
• vSAN ESA drives average of 17%
higher CPU usage
• CPU bottleneck being experienced

workload with vSAN ESA resulting in high

latency at high IOPS

RAID 5
70% reads / 30% writes
Space efficiency:
• vSAN OSA –
Compression &
deduplication
• vSAN ESA –
Compression only

Internal Use - Confidential 157 Copyright © Dell Inc. All Rights Reserved.
 Read I/Os OSA performed better than
Peak ESA, averaging 10% higher
throughput overall in these

performan random read tests.

ce by I/O
size
”xSizes” test
RAID 5
Space efficiency:
• vSAN OSA –
Compression &
deduplication
• vSAN ESA –
Compression only

Internal Use - Confidential 158 Copyright © Dell Inc. All Rights Reserved.
 Write I/Os • ESA performed significantly better

Peak than OSA in these random write

tests

performan
• Averaging 2x throughput overall
• Substantial increase in
throughput for the 512Kb I/O
ce by I/O with 3x throughput compared to
OSA

size
”xSizes” test
RAID 5

Space efficiency:
• vSAN OSA –
Compression &
deduplication
• vSAN ESA –
Compression only

Internal Use - Confidential 159 Copyright © Dell Inc. All Rights Reserved.
 Large I/O
max
sustained
sequential
Writes OSA destaging
After OSA destaging, the
throughput for OSA drops
significantly compared to
ESA, with average of 4x
“Steady state” test more throughput with ESA

RAID 5

Space efficiency:
• vSAN OSA –
Compression &
deduplication
• vSAN ESA –
Compression only
Internal Use - Confidential 160 Copyright © Dell Inc. All Rights Reserved.
 Key • ESA outperforms OSA for small and medium I/O mixed workloads
Takeaways • Similar performance for large block I/O mixed workloads for low IOPS, with
for VD-4000 increase in latency and decrease in IOPS seen for higher IOPS for ESA

with VxRail • CPU usage pushed to limit with ESA signifying bottleneck for higher
IOPS driving latency and IOPS difference
8.0.210
• ESA performed significantly better than OSA for random write peak
performance test

• 2x higher throughput with ESA

• 3x higher throughput in the 512K test

• For sustained sequential writes at large I/O block size – ESA performed
significantly better than OSA with an average of 4x higher throughput

Internal Use - Confidential 161 Copyright © Dell Inc. All Rights Reserved.
 16G VxRail platforms
Performance of 16G VxRail with Sapphire Rapids

Internal Use - Confidential 162 Copyright © Dell Inc. All Rights Reserved.
 VxRail 16th Generation platforms

vSAN ESA
Advantage
• Core Features of VxRail 16th Generation with Sapphire Rapids

1 2 3 4
Choice of vSAN More CPU & More Storage Built-in
• vSAN OSA Increased and More Accelerator for AI
or
Bandwidth CPU Memory • Choice of Intel 4th
• Intel 4 Generation
th
• Up to 368TB of Gen Xeon Scalable
• vSAN ESA
Sapphire Rapids storage processor comes
• Single Tier CPU with built-in Intel®
architecture • Up to 8TB of RAM AMX accelerator
• Up to 56 cores per socket improves AI
• Lower TCO per/CPU
• 8 Memory Channels Inference
• More Usable • PCIe Gen 5 Workloads
storage • DDR5 4800 MT/s

• RAID 1
performance with
RAID 5/6 capacity
Internal Use - Confidential 163 Copyright © Dell Inc. All Rights Reserved.
 1 6 T H G E N E R AT I O N P E R F O R M A N C E

VMmark comparison with vSAN

OSA
15th Generation Ice Lake versus 16th Generation Sapphire
Rapids

Internal Use - Confidential 164 Copyright © Dell Inc. All Rights Reserved.
 16th Generation platforms with vSAN OSA
More workloads per server The numbers
VMmar k S c o r es wit h v S A N OS A • VMmark 3.1.1 Scores
(h ig h er is b ett er )
45
• 16th Generation PowerEdge with
40 39.11
Sapphire Rapids (56C) can hold
35

30
1.6X more workloads compared with
25
24.48 a similar 15G cluster with Ice Lake
20 CPU (40c)
15

10
• Footprint reduction and workload
5 consolidation
0
4x PowerEdge 15th Gen. Ice Lake @26 4x PowerEdge 16th Gen. Sapphire
Tiles [494 VM's] Rapids @42 Tiles (798 VMs)

Internal Use - Confidential 165 Copyright © Dell Inc. All Rights Reserved.
 V X R A I L 1 6 T H G E N E R AT I O N P E R F O R M A N C E

New Adaptive Write Path

with vSAN ESA 8.0 U1
VxRail Generation Comparisons

Internal Use - Confidential 166 Copyright © Dell Inc. All Rights Reserved.
 vSAN ESA – New Adaptive Write Path
• New with vSAN ESA 8.0 U1
• Alternative write path for large I/O Block sizes
• Dynamically uses alternate write path for guest VM writes
using large I/O sizes
• Default write path
• Large I/O write path
• Large writes bypass Durable log of vSAN ESA’s Log-
structured File System
• Commits I/Os as a full-stripe write
• Metadata is written to the Durable Log
• Improves performance of streaming writes with no additional
complexity
• Reduced write amplification and CPU utilization
• Higher throughput and lower latency for sequential write
workloads
Internal Use - Confidential 167 Copyright © Dell Inc. All Rights Reserved.
 VxRail vSAN ESA Comparison:
Some differences between the systems affect performance, especially vSAN version & CPU type

16th Generation 15th Generation

VxRail 8.0.120 VxRail 8.0.000

• 6 x VxRail VP-760 • 6 x VxRail P670N

• 2 x Intel(R) Xeon(R) • 2 x Intel(R) Xeon(R)
Gold 6430 CPU @ Gold 6338 (4 nodes) &
2.10GHz 6330 (2 nodes) CPU
 64 cores per node both @ 2.00 GHz
 6338 - 64 cores per node
• 1,024 GB RAM  6330 – 56 cores per node
• Broadcom BCM57414 • 1,024 GB RAM
NetXtreme-E 25GbE
• Intel E810-XXV 25GbE
• 12 NVMe drives per host
 Intel P5620 MU 3.2TB
• 11 NVMe drives per host
 Intel P5600 MU 3.2TB
• vSAN ESA 8.0.u1
• vSAN ESA 8.0

Internal Use - Confidential 168 Copyright © Dell Inc. All Rights Reserved.
 Peak performance by I/O size
AKA “xSizes” tests, 6 node, storage policy = RAID-5, space efficiency = compression

VxRail P670N 8.0.000 vs VP-760 8.0.120 VxRail P670N 8.0.000 vs VP-760 8.0.120
 P670N, 8.0.000  VP760, 8.0.120  P670N, 8.0.000  VP760, 8.0.120
11,000 12,000
10,000 11,000

9,000 10,000 + 105%

+ 95%
9,000
8,000
Max Throughput MB/s

Max Throughput MB/s

8,000
7,000
+ 19% 7,000 + 23 %
6,000
+ 18% 6,000 + 21 %
5,000
5,000
4,000
4,000
3,000 3,000
2,000 2,000
1,000 1,000
0 0
4K 8K 16K 32K 64K 128K 256K 512K 4K 8K 16K 32K 64K 128K 256K 512K
I/O Size I/O Size

Random Write Sequential Write

Internal Use - Confidential 169 Copyright © Dell Inc. All Rights Reserved.
 Peak performance by I/O size
6 nodes, Storage policy = RAID-5, space efficiency = compression

VxRail P670N 8.0.000 vs VP-760 8.0.120 VxRail P670N 8.0.000 vs VP-760 8.0.120
 P670N, 8.0.000  VP760, 8.0.120  P670N, 8.0.000  VP760, 8.0.120
11,000 12,000
10,000 11,000
Big improvement with 2x more sequential
9,000 vSAN ESA 8.0 U1 for +95% 10,000 write throughput with +105%
larger I/O block size 9,000
vSAN ESA 8.0U1
8,000
Max Throughput MB/s

Max Throughput MB/s

7,000 8,000
7,000 +21% +23%
6,000 +18% +19%
6,000
5,000
5,000
4,000
4,000
3,000
3,000
2,000
2,000
1,000 1,000
0 0
128K 256K 512K 128K 256K 512K
I/O Size I/O Size

Random Write Sequential Write

Internal Use - Confidential 170 Copyright © Dell Inc. All Rights Reserved.
 VxRail vSAN OSA R1 versus vSAN ESA R5 Comparison
Some differences between the systems affect performance

16th Generation 14th Generation

VxRail 8.0.120 VxRail 7.0.350

• 6 x VxRail VP-760 • 6 x VxRail P570F

• 2 x Intel(R) Xeon(R) • 2 x Intel(R) Xeon(R)
Gold 6430 CPU @ Gold 8168CPU @ 3.1
2.10GHz GHz
  48 cores per node
64 cores per node

• 1,024 GB RAM • 384 GB RAM

• Broadcom BCM57414 • Broadcom Adv. 25 GbE
NetXtreme-E 25GbE • 2 Disk Groups per host:
• 12 NVMe drives per host  1 x Cache 0.8 TB
 Intel P5620 MU 3.2TB  3 x 1.78 SAS TB

• vSAN ESA 8.0.u1 • vSAN OSA 6.7 U3

Internal Use - Confidential 171 Copyright © Dell Inc. All Rights Reserved.
 Peak performance by I/O size
16G vSAN ESA RAID 5 (with compression) versus 14G vSAN OSA RAID 1 (no space efficiency)

VxRail P570F OSA 6.7u3 RAID 1 vs VP-760 ESA 8.0u1 RAID 5 VxRail P570F OSA 6.7u3 RAID 1 vs VP-760 ESA 8.0u1 RAID 5
 P570F, 7.0.350  VP760, 8.0.120  P570F, 7.0.350  VP760, 8.0.120
11,000 12,000
10,000 4.8x more random +4.8x 11,000 4.4x more sequential +4.4x
write throughput with write throughput with
9,000 10,000
VxRail 16G and VxRail 16G and
vSAN ESA 8.0U1 9,000 vSAN ESA 8.0U1
Max Throughput MB/s

8,000

Max Throughput MB/s

7,000 8,000

+2.8x 7,000 +2.7x +2.7x

6,000 +2.8x
6,000
5,000
5,000
4,000
4,000
3,000
3,000
2,000
2,000
1,000 1,000
0 0
128K 256K 512K 128K 256K 512K
I/O Size I/O Size
Random Write Sequential Write

Internal Use - Confidential Copyright©©Dell

172Copyright
172 DellInc.
Inc.
AllAll Rights
Rights Reserved.
Reserved.
 V X R A I L 1 6 T H G E N E R AT I O N P E R F O R M A N C E

VxRail All-NVMe with vSAN ESA:

Storage policy comparisons

RAID 5 versus RAID 6

Internal Use - Confidential 173 Copyright © Dell Inc. All Rights Reserved.
VxRail
8.0.120

• 6 x VxRail VP-760
• 2 x Intel(R) Xeon(R) Gold
6430 CPU @ 2.10GHz
 64 Cores per node

• 1,024 GB RAM
• Broadcom BCM57414
NetXtreme-E 25GbE
• 12 NVMe drives per host
• Intel P5620 MU 3.2TB
• vSAN ESA 8.0.u1

Internal Use - Confidential 174 Copyright © Dell Inc. All Rights Reserved.
174
 VxRail VP-760 8.0.120, RAID-5 vs RAID-6
OLTP 8K, vSAN ESA, 6 Nodes, 12 Drives/Node

Small I/O 1.6

mixed 1.4
RAID5
workload 1.2
RAID6

Response Time (ms)

1.0
70% reads
30% writes 0.8
Average
space efficiency = + 15% IOPS
compression 0.6
- 2% response
time
0.4

0.2
RAID-6, Compression RAID-5, Compression
0.0
0 200,000 400,000 600,000 800,000 1,000,000

IOPS

Internal Use - Confidential 175 Copyright © Dell Inc. All Rights Reserved.
 VxRail VP-760 8.0.120, RAID-5 vs RAID-6
OLTP 32K, vSAN ESA, 6 Nodes, 12 Drives/Node

Medium 3.0

I/O mixed
workload
2.5

Response Time (ms)

2.0
70% reads
RAID6 RAID5
30% writes
1.5
space efficiency =
compression
1.0 Average
+ 18% IOPS
- 4% response
0.5 time
RAID-6, Compression RAID-5, Compression

0.0
0 50,000 100,000 150,000 200,000 250,000 300,000

IOPS

Internal Use - Confidential 176 Copyright © Dell Inc. All Rights Reserved.
 VxRail VP-760 8.0.120, RAID-5 vs RAID-6
OLTP 64K, vSAN ESA, 6 Nodes, 12 Drives/Node

Large I/O 5.0

mixed 4.5

workload 4.0
RAID6 RAID5
3.5

Response Time (ms)

70% reads
3.0
30% writes
2.5
space efficiency =
compression 2.0

1.5
Average
1.0 + 24% IOPS
- 5% response time
0.5
RAID-6, Compression RAID-5, Compression
0.0
0 25,000 50,000 75,000 100,000 125,000 150,000 175,000

IOPS

Internal Use - Confidential 177 Copyright © Dell Inc. All Rights Reserved.
 VxRail VP-760 8.0.120, RAID-5 vs RAID-6
RW xSizes, vSAN ESA, 6 Nodes, 12 Drives/Node

Peak 11,000

performan 10,000

Average all tests

ce by I/O 9,000
+ 34% throughput

size 8,000

Max Throughput MB/s

7,000

Random 6,000

Writes 5,000

Maximum throughput 4,000

MB/s
3,000
Space efficiency default
2,000
= compression
1,000

0
4K 8K 16K 32K 64K 128K 256K 512K

RAID-6, Compression I/O Size RAID-5, Compression

Internal Use - Confidential 178 Copyright © Dell Inc. All Rights Reserved.
 VxRail VP-760 8.0.120, RAID-5 vs RAID-6
SW xSizes, vSAN ESA, 6 Nodes, 12 Drives/Node

Peak 12,000

performan
11,000

10,000 Average all tests

ce by I/O 9,000
+ 34% throughput

size

Max Throughput MB/s

8,000

Sequential 7,000

writes 6,000

5,000

Maximum throughput MB/s 4,000

3,000
Space efficiency default =
compression 2,000

1,000

0
4K 8K 16K 32K 64K 128K 256K 512K

RAID-6, Compression I/O Size RAID-5, Compression

Internal Use - Confidential 179 Copyright © Dell Inc. All Rights Reserved.
 VxRail VP-760 8.0.120, RAID-5 vs RAID-6
SW 64K Max, vSAN ESA, 6 Nodes, 12 Drives/Node

Large I/O 6,000

RAID5
max 5,500

sustained 5,000
Average + 37%
sequential
4,500 throughput
RAID6
4,000
writes
Throughput MB/s
3,500

3,000
AKA “steady state”
tests, 2,500

space efficiency = 2,000

compression 1,500

1,000

Throughput MB/s 500

RAID-6, Compression RAID-5, Compression
0
0 10 20 30 40 50 60

Elapsed Time - Minutes

Internal Use - Confidential 180 Copyright © Dell Inc. All Rights Reserved.
 VxRail VP-760 8.0.120, RAID-5 vs RAID-6
SW 64K Max, vSAN ESA, 6 Nodes, 12 Drives/Node

Large I/O 7.0

max
6.5
RAID6
6.0

sustained 5.5

sequential 5.0
RAID5

Response Time (ms)

writes
4.5

4.0

3.5 Average 27%

AKA “steady state” lower response
tests, 3.0 time
2.5
space efficiency = 2.0
compression
1.5

1.0
Response Time 0.5
(ms) RAID-6, Compression RAID-5, Compression
0.0
0 10 20 30 40 50 60

Elapsed Time - Minutes

Internal Use - Confidential 181 Copyright © Dell Inc. All Rights Reserved.
 • For PowerEdge and VxRail with vSAN OSA, 16G with
Key Sapphire Rapids (56c) can hold 1.6X more workloads
Takeaways compared with 15G with Ice Lake CPU (40c)
for 16th • New adaptive write path with vSAN ESA 8.0 U1 provides
Generation much greater performance for large write I/Os when
VxRail compared to vSAN ESA 8.0
o Up to 95% more throughput for random writes
o Up to 105% more throughput for sequential writes

• RAID-5 consistently outperforms RAID-6 with vSAN ESA

o More IOPS and lower latency for small, medium and large
mixed workloads
o On average, 37% more throughput and 27% lower latency
for large I/O sustained sequential writes with RAID-5

Internal Use - Confidential 182 Copyright © Dell Inc. All Rights Reserved.
 No
v em
be
r2
02
3

AI Benchmarking
with VxRail
Performance of 16G VxRail with Intel® AMX
accelerator

Internal Use - Confidential 183 Copyright © Dell Inc. All Rights Reserved.
VxRail 8.X VxRail VE-660

• 4 x VxRail VE-660

• 2 x Intel(R) Xeon(R) Gold

6430
• 512 GB RAM
• Broadcom Adv. Dual 25Gb
Ethernet
• 2 DG per host:
 1 x Cache, 4 x Capacity,

• vSAN OSA 8.0 U2

Internal Use - Confidential 184 Copyright © Dell Inc. All Rights Reserved.
 Resnet50 Image classifi cati on performance on
benchmark Tensorfl ow 2.11 using ResNet50
results 9000 8559
3.1x faster
• Inference performance 8000
with AMX
increased by 3.1x for 7000
int8 when compared
with previous VxRail 6000
generation
5000

images/sec
• int8 offers performance 4000
with minimal impact to 2941
3000 2739
accuracy
2000
• This is where AMX
proved to be very 1000

efficient
0
int8

VxRail 15G 6330(28c) CPU VxRail 15G 6338(32c) CPU VxRail 16G 6430(32c) CPU

Internal Use - Confidential 185 Copyright © Dell Inc. All Rights Reserved.
 BERT NLP on TensorFlow 2.11 using BERT
benchmark VxRail 15G vs VxRail 16G with Intel®
results AMX
4.0
• The performance 3.70
3.7x faster
achieved on BERT with 3.5 with AMX
VxRail 16G and Intel ®
3.0
AMX is 3.7 times of
what we achieved on

normalized performance
2.5
the previous VxRail
15G gen. 2.0

1.5

1
1.0

0.5

0.0
int8

VxRail 15G 6338(32c) CPU VxRail 16G 6430(32c) CPU

Internal Use - Confidential 186 Copyright © Dell Inc. All Rights Reserved.
 Key
Takeaways
for AI – VxRail 16G is now more ready than ever for AI
Benchmarki workloads with Intel® AMX
ng – VxRail VE-660 nodes with built-in Intel® AMX improves
AI performance:
▪ 3.1x for Image Classification
▪ 3.7x for Natural Language Processing (NLP)
– Intel® AMX delivers a cost-effective way to run AI
workloads without the need of a dedicated GPU

Internal Use - Confidential 187 Copyright © Dell Inc. All Rights Reserved.
 VXRAIL VD-4000 PERFORMANCE

VxRail VD-4000
Performance Overview
Internal Use - Confidential 188 Copyright © Dell Inc. All Rights Reserved.
 VxRail VxRail VD-4000 using VD-4520c nodes
7.0.420
VD-4520c
• 6 x VxRail VD-4520c
• 1 x Intel(R) Xeon(R)
D-2776NT CPU @ 2.10 GHz
 16 Cores

• 128 GB RAM
• Intel(R) E823-C 25 GbE
• 1 DG per host:
 1 x Cache, Micron 7400
NVMe MU M.2 - 800 GB
 3 x Capacity, Micron 7400
NVMe RI M.2 - 1.92 TB

• vSAN OSA 7.0.U3

Internal Use - Confidential 189 Copyright © Dell Inc. All Rights Reserved.
189
 VxRail VD-4520c, 7.0.420, 6 Nodes, All NVMe
OLTP 4K, SP Comparison, no SE, 1 DG

Small I/O 3.5

Mixed 3.0

Workload
2.5

Response Time (ms)

70% random reads RAID6 RAID5 RAID1

30% random writes 2.0

Space efficiency = none

1.5

1.0

0.5

RAID1FTT1 RAID5 RAID6

0.0
0 50,000 100,000 150,000 200,000 250,000

IOPS

Internal Use - Confidential 190 Copyright © Dell Inc. All Rights Reserved.
 VxRail VD-4520c, 7.0.420, 6 Nodes, All NVMe
OLTP 16K, SP Comparison, no SE, 1 DG

Medium 4.0

I/O Mixed 3.5

Workload 3.0

RAID6 RAID5 RAID1

Response Time (ms)

70% random reads 2.5

30% random writes

2.0
Space efficiency = none

1.5

1.0

0.5
RAID1FTT1 RAID5 RAID6
0.0
0 50,000 100,000 150,000 200,000 250,000

IOPS

Internal Use - Confidential 191 Copyright © Dell Inc. All Rights Reserved.
 VxRail VD-4520c, 7.0.420, 6 Nodes, All NVMe
OLTP 64K, SP Comparison, no SE, 1 DG

Large I/O 8

Mixed 7

Workload 6

Response Time (ms)

70% random reads 5
30% random writes
4
Space efficiency = none

3
RAID6 RAID5 RAID1

1
RAID1FTT1 RAID5 RAID6
0
0 10,000 20,000 30,000 40,000 50,000 60,000 70,000 80,000

IOPS

Internal Use - Confidential 192 Copyright © Dell Inc. All Rights Reserved.
 General VD-
4000
performance
takeaways

Similar RAID and • Similar performance differences of Raid 1, 5, 6 compared to

Storage Efficiency
characterizations as other VxRail node types
other OSA based nodes
• Not shown is similar delta of other OSA nodes using storage
efficiency (SE)

Internal Use - Confidential 193 Copyright © Dell Inc. All Rights Reserved.
 VXRAIL VD-4000 PERFORMANCE

Comparative platform
test-
P570 single socket to VD-4000

Internal Use - Confidential 194 Copyright © Dell Inc. All Rights Reserved.
 VxRail VD-4000 using VD-4520c nodes
7.0.420

• 4 x VxRail VD-4520c
• 1 x Intel(R) Xeon(R)
D-2776NT CPU @ 2.10 GHz
 16 Cores

• 128 GB RAM
• Intel(R) E823-C 25 GbE
• 2 DGs per host:
 1 x Cache, Micron 7400 NVMe
MU M.2 - 800 GB
 3 x Capacity, Micron 7400
NVMe RI M.2 - 1.92 TB

• vSAN OSA 7.0.U3

Internal Use - Confidential 195 Copyright © Dell Inc. All Rights Reserved.
 VxRail VxRail P670F
8.0.000

• 4 x VxRail P670F
• 1 x Intel(R) Xeon(R) Gold
6330 CPU @ 2.00GHz
 28 Cores

• 512 GB RAM
• Broadcom Adv. Dual 25 GbE
• 2 DGs per host:
 1 x Cache, Intel NVMe P5600
MU U.2 1.6 TB
 4 or 5 x Capacity, Kioxia SAS
PM6 RI U.2 1.92 TB

• vSAN OSA 8.0

Internal Use - Confidential 196 Copyright © Dell Inc. All Rights Reserved.
 VxRail 1S CPU, VD-4520c vs P670F
Small Block Random I/Os - Cache Drives
1,100,000
VD-4520c 1,000,000
VD-4520c Micron 7400 (M.2) NVMe - MU P670F Intel P5600 NVMe - MU

vs P670F
900,000
Cache

Manufacturer Reported IOPS

Drives 800,000

700,000
Small random I/Os
600,000

Manufacturer reported 500,000 1,000,000

IOPS
400,000

300,000

200,000

230,000 260,000
100,000
118,000
0
RR 4K IOPS RW 4K IOPS

I/O Type & Size

Internal Use - Confidential 197 Copyright © Dell Inc. All Rights Reserved.
 VxRail 1S CPU, VD-4520c vs P670F
Large Block Sequential I/Os - Cache Drives
8,000
VD-4520c VD-4520c Micron 7400 (M.2) NVMe - MU P670F Intel P5600 NVMe - MU

vs P670F 7,000

Cache

Manufacturer Reported MB/S

6,000
Drives
Large sequential I/Os 5,000

4,000
Manufacturer reported 7,000
throughput MB/s
3,000

4,400 4,300
2,000

1,000
1,000
0
SR 128K MB/s SW 128K MB/s

I/O Type & Size

Internal Use - Confidential 198 Copyright © Dell Inc. All Rights Reserved.
 VXRAIL VD-4000 PERFORMANCE

Peak Performance
with CPU and power usage comparisons

Internal Use - Confidential 199 Copyright © Dell Inc. All Rights Reserved.
 VxRail 1S CPU, VD-4520c vs P670F
RW xSizes, 4 Nodes, 2 DGs, RAID5 no SE
3,500
Peak VD-4520c 7.0.420 P670F 8.0.000

Performan 3,000

ce
2,500

Max Throughput MB/s

Random writes
2,000

Maximum throughput MB/s

1,500

1,000

500

0
4K 8K 16K 32K 64K 128K 256K 512K

I/O Size

Internal Use - Confidential 200 Copyright © Dell Inc. All Rights Reserved.
 VxRail 1S CPU, VD-4520c vs P670F
RW xSizes, 4 Nodes, 2 DGs, RAID5 no SE

Peak
100

Performan 90

ce

Average CPU Usage % for All Hosts

80

70
VD-4520c vs P670F
Random writes Avg + 65% CPU usage
60

50
CPU usage percent
40
Average for all hosts
30

20

10
VD-4520c 7.0.420 P670F 8.0.000
0
0 10 20 30 40 50 60 70 80 90 100

Collection Interval (~ 20 Seconds Each)

Internal Use - Confidential 201 Copyright © Dell Inc. All Rights Reserved.
 VxRail 1S CPU, VD-4520c vs P670F
RW xSizes, 4 Nodes, 2 DGs, RAID5 no SE

Peak
350

Performan 300

ce

Average Power Watts for All Hosts

250
P670F vs VD-4520c
Avg + 59% power usage
Random writes
200

Power usage in watts

150
Average for all hosts

100

50

VD-4520c 7.0.420 P570F 8.0.000

0
0 10 20 30 40 50 60 70 80 90 100

Collection Interval (~ 20 Seconds Each)

Internal Use - Confidential 202 Copyright © Dell Inc. All Rights Reserved.
 V X R A I L J A F FA P E R F O R M A N C E – P H A S E 2

Mixed Workloads
with new IOPS per watt metric

Internal Use - Confidential 203 Copyright © Dell Inc. All Rights Reserved.
 VxRail 1S CPU, VD-4520c vs P670F
OLTP 8K, 4 Nodes, 2 DGs, RAID5 no SE

Small I/O 1.6

Mixed 1.4

Workload 1.2
VD-4520c

P670F
70% random reads

Response Time (ms)

1.0

30% random writes

0.8

Storage policy = RAID5 0.6

Space efficiency = none

0.4

0.2
VD-4520c 7.0.420 P670F 8.0.000
0.0
0 50,000 100,000 150,000 200,000 250,000 300,000

IOPS

Internal Use - Confidential 204 Copyright © Dell Inc. All Rights Reserved.
 VxRail 1S CPU, VD-4520c vs P670F
OLTP 8K at Max IOPS, 4 Nodes, 2 DGs, RAID5 no SE

Small I/O 1,400

Average IOPS per Watt of Power (at Max IOPS)

Mixed 1,200
IOPS/Watt - 1,119
Workload
1,000

Average IOPS
Per watt of power used 800
IOPS/Watt - 697

600
Calculated at maximum
IOPS from ramp-up test 400

VD-4520c has 61% more 200

IOPS per watt than
P670F with OLTP 8K at
max IOPS 0
VD-4520c 7.0.420 P670F 8.0.000

Internal Use - Confidential 205 Copyright © Dell Inc. All Rights Reserved.
 VxRail 1S CPU, VD-4520c vs P670F
OLTP 64K, 4 Nodes, 2 DGs, RAID5 no SE

Large I/O 2.6

Mixed
2.4

2.2

Workload 2.0
VD-
4520c
1.8

Response Time (ms)

70% random reads 1.6
30% random writes P670F
1.4

1.2

Storage policy = RAID5 1.0

Space efficiency = none 0.8

0.6

0.4

0.2
VD-4520c 7.0.420 P670F 8.0.000
0.0
0 20,000 40,000 60,000 80,000 100,000 120,000

IOPS

Internal Use - Confidential 206 Copyright © Dell Inc. All Rights Reserved.
 VxRail 1S CPU, VD-4520c vs P670F
OLTP 64K at Max IOPS, 4 Nodes, 2 DGs, RAID5 no SE

Large I/O 500

IOPS/Watt - 460

Average IOPS per Watt of Power (at Max IOPS)

Mixed 450

Workload 400
IOPS/Watt - 357
350
Average IOPS
300
Per watt of power used
250

Calculated at maximum 200

IOPS from ramp-up test 150

100
VD-4520c has 29% more
IOPS per watt than 50
P670F with OLTP 64K at
max IOPS 0
VD-4520c 7.0.420 P670F 8.0.000

Internal Use - Confidential 207 Copyright © Dell Inc. All Rights Reserved.
 Key
Takeaways
for VD- – Single socket VxRail model P670F mainly outperforms VD-4520c, based on
4000 traditional performance metrics like IOPS & latency
– Main reason is lesser performance capabilities of VD-4000 hardware,
especially of Micron 7400 M.2 vs Intel P5600 U.2 NVMe MU cache drives.
– VD-4520c uses significantly less power than P670F.
– And using a new measurement, VD-4520c outperforms P670F - up to
61% higher IOPS per watt of power.

Internal Use - Confidential 208 Copyright © Dell Inc. All Rights Reserved.
 Performance overview

VxRail with vSAN 8.0

ESA
Internal Use - Confidential 209 Copyright © Dell Inc. All Rights Reserved.
 VxRail with VSAN ESA

Performance Core features

1 2 3 4
Single Tier Optimized for More Resilient Ready for the
Architecture performance most demanding
workloads
• Big Data, Decision
• More usable • Optimized and Support Systems,
capacity only available for • Fault domain is Databases
• Less complexity All NVMe VxRail now limited to
platforms single drives
P670N/E660N instead of disk
• Efficient data groups
services • Faster vSAN
• RAID5/6 delivers resyncs
Internal Use - Confidential performance now
210 Copyright © Dell Inc. All Rights Reserved.
 VxRail ESA
VxRail platforms tested

VXRAIL 8.0.00 ESA VXRAIL 8.0.00 OSA

• 6 x VxRail P670N • 6 x VxRail P670N

• 2x Intel(R) Xeon(R) Gold • 2x Intel(R) Xeon(R) Gold
6330 CPU @ 2.00GHz 6330 CPU @ 2.00GHz
• 28 Cores • 28 Cores

• 1024 GB RAM • 1024 GB RAM

• 2x25GbE (LACP) for • 2x25GbE (LACP) for
VSAN VSAN
• 11x NVMe drives/host: • 4 DG’s per host:
• Intel P5600 3.2TB • 1 Cache Intel P5600 3.2TB
• 2x Capacity Intel P5600 3.2TB
• VxRail version 8.0.000
• VxRail version 8.0.000

These 2 systems are nearly identical the same:

We just deployed different VSAN versions & OSA
had one additional capacity drive.
Internal Use - Confidential 211 Copyright © Dell Inc. All Rights Reserved.
 Penalty Free Compression in ESA
<5% at peak and nearly identical performance up to 50% load

vSAN ESA Compression vs no Compression

70/30 32KB RAID5
2.50 2.24

2.00
2.11

1.50
latency (ms)

1.00

0.50

0.00
0 50000 100000 150000 200000 250000
IOPS

No compression Compression

Internal Use - Confidential 212 Copyright © Dell Inc. All Rights Reserved.
 ESA vs OSA
Mixed workloads with RAID5/RAID6

70/30 8KB RAID5 50GbE 70/30 8KB RAID6 50GbE

ESA vs OSA ESA vs OSA
2.50 2.50
1.98

2.00 2.00
1.33
1.56
1.15
1.50 1.50

latency (ms)
latency (ms)

1.00 1.00

0.50 0.50
+49% more IOPS
+25% more IOPS 31% Lower latency
0.00 39% Lower latency 0.00
0 00 00 00 00 00 00 00 00 00 00 0 00 00 00 00 00 00 00 00 00 00
000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
IOPS IOPS

ESA OSA ESA OSA

Internal Use - Confidential 213 Copyright © Dell Inc. All Rights Reserved.
 ESA vs OSA
Peak Performance with RAID6

P e ak P e r fo r m a n c e E S A v s O S A P e ak P e r fo r m a n c e E S A v s O S A
4 KB / RA ID6 / 5 0 GbE 6 4 KB / RA ID6 / 5 0 GbE

1200000 8000
7508
1095778
7000
1000000 6453
2.2X or
3.5X or +119%
864593 6000
+250%
5290
800000 +22%
5000
2.4X or
+140%
600000 4000

MB/s
IOPS

3434

3000
400000 358159
306933
2000

200000
1000

0 0
Rand. Write Seq. Write Rand. Write Seq. Write

OSA ESA OSA ESA

Internal Use - Confidential 214 Copyright © Dell Inc. All Rights Reserved.
 ESA vs OSA Throughput Comparison
Sustained write performance with RAID5 is much advantage with vSAN ESA
100% Seq. Write 64KB during 1 hour
RAID5
11000 50% drop

10000 9635

9000

8000
recovers
+87% throughput
MB/s

7000

6000
never
5154
recovers
5000

4000

3000
0 500 1000 1500 2000 2500 3000 3500
Elapsed times (seconds)

ESA OSA

Internal Use - Confidential 215 Copyright © Dell Inc. All Rights Reserved.
 VxRail vSAN RAID policy comparison OSA vs ESA
No Storage Efficiency OSA; Compression Enabled in ESA

6xVxRail E660F 7.0.320 70/30 8KB

RAID 6 2.3X (134%)
1.80
RAID 5 2.3X (127%)
1.60 RAID 1 77%
1.40
latency (ms)

1.20

1.00

0.80

0.60

0.40
0 200000 400000 600000 800000 1000000 1200000
IOPS

OSA7-RAID1 OSA7-RAID5 OSA7-RAID6 ESA8-RAID1 ESA8-RAID5 ESA8-RAID6

Internal Use - Confidential 216 Copyright © Dell Inc. All Rights Reserved.
 VxRail vSAN RAID policy comparison OSA vs ESA
No Storage Efficiency OSA RAID 1; Compression in ESA RAID 5

6xVxRail E660F 7.0.320 70/30 8KB

1.80

1.60

1.40
+ 57% IOPS
latency (ms)

1.20

1.00

0.80

0.60

0.40
0 200000 400000 600000 800000 1000000 1200000
IOPS

OSA7-RAID1 ESA8-RAID5

Internal Use - Confidential 217 Copyright © Dell Inc. All Rights Reserved.
 VxRail vSAN RAID policy comparison OSA vs ESA
No Storage Efficiency OSA RAID 1; Compression in ESA RAID 6

6xVxRail E660F 7.0.320 70/30 8KB

1.60
33% more IOPS 33% More IOPS
1.40 AND
33% more usable capacity (TBu)
1.20
latency (ms)

1.00

0.80

0.60

0.40
0 200000 400000 600000 800000 1000000 1200000
IOPS

OSA7-RAID1 ESA8-RAID6

Internal Use - Confidential 218 Copyright © Dell Inc. All Rights Reserved.
 Takeaways

• RAID5/6 now delivers performance

and reduces $/GB- USE THEM in
vSAN ESA configurations!

• ESA doesn't deliver full gains with

25GbE. Go for 100GbE

• Compression is nearly penalty free

• Throughput is significantly better

for sustained writes in ESA vs
OSA

Internal Use - Confidential 219 Copyright © Dell Inc. All Rights Reserved.
 Ju
ne
20
23

TPC-DS Analysis
Choosing 100 Gbit/sec vs 25 Gbit/sec on VxRail with VSAN ESA

Note: There will not be an external TPC-DS benchmark and these results should not be used for
any public claims or statements.
 Goals

What Why How

• Measure decision support • Highlight an important • Leveraging the TPC-DS

workloads performance enterprise workload that benchmark
when running on VSAN can benefit from 100GbE • Collecting several workload
ESA with 25GbE vs for ESA related performance
100GbE
• Show the performance metrics
benefits for workloads
running on 100GbE

Internal Use - Confidential 221 Copyright © Dell Inc. All Rights Reserved.
 VxRail
Platforms
Tested VXRAIL 8.0.0 ESA VXRAIL 8.0.0 ESA

• 6 x VxRail P670N • 6 x VxRail P670N

• 2x Intel(R) Xeon(R) Gold • 2x Intel(R) Xeon(R) Gold
6338 CPU @ 2.00GHz 6338 CPU @ 2.00GHz
• 32 Cores • 32 Cores

• 1TB RAM • 1TB RAM

• 2 x Broadcom NetXtreme • 2 x Broadcom NetXtreme-
E-Series Quad-port 25Gb E P2100D BCM57508
100Gb
• 12 X Dell Ent NVMe
P5600_MU_U.2 1.6TB • 12 X Dell Ent NVMe
Disks P5600_MU_U.2 1.6TB
Disks

These 2 systems are exactly the same:

The only difference is the 25Gb/sec vs the
100Gb/sec nics
Internal Use - Confidential 222 Copyright © Dell Inc. All Rights Reserved.
 TPC-DS VM
Specs

TPC-DS VM

VM
• CPU: 18
• Memory 256 GB
• Disk OS: 25 GB
• Disk Data: 3.5 TB
• Nics: E1000
• OS: Debian Linux 10
(Turnkey Linux)
• Postgres: (PostgreSQL)
11.17
database

Internal Use - Confidential 223 Copyright © Dell Inc. All Rights Reserved.
 Support)

TPC-DS is a decision support benchmark that models several generally applicable aspects of a
decision support system, including queries and data maintenance. The benchmark provides a
representative evaluation of performance as a general-purpose decision support system. The
imaginary retailer sells its goods via three different distribution channels:

• Store
• Catalog
• Internet
Testing Sections
#1 Data Generation

dsdgen scripts used to

generate csv style data files

Time to create 1TB

of data is the measure

#2 Data Load
#3 Data Query

csv style files from Stage 1 Set of Queries run on the

loaded to Postgres DB 1TB DB

Average and Total query

Time to load 1TB time the measure
of data into the DB
 1 – Data
Generation

• Ultimately there is no noticeable difference between 25G

and 100G < 1%
• Reason being that CPU contention is always met before
network/storage contention

Internal Use - Confidential 226 Copyright © Dell Inc. All Rights Reserved.
 2 - Data Load • psql is used to load the files from Stage 1 into the Postgres DB.
Overview • This Test can again be run multi-threaded in our case we ran 16
threads
 Why 16?
– Was easier to keep to 16 as that was used in Stage 1 which meant
This is the second step there was a pattern to the created files which could be copies to
in the workflow
scripts –
• What did we measure?
1. Throughput while loading
2. Total Time to load the data

Internal Use - Confidential 227 Copyright © Dell Inc. All Rights Reserved.
 2 - Data Load Max Throughput 18.94 GB/s
100GbE vSAN Throughput Loading
Throughput
20
15
10
5
Throughout a lot higher 0
[Link] AM[Link] PM[Link] PM[Link] PM[Link] PM[Link] AM[Link] AM[Link] AM[Link] AM
on the 100GbE system
Total GB/s Read-GB/s Write-GB/s
There is 47.05% more
throughput on the 100G Max Throughput 12.88 GB/s
System 25GbE vSAN Throughput Loading
15

10

0
[Link] AM[Link] PM[Link] PM[Link] AM[Link] AM[Link] AM[Link] PM[Link] PM[Link] AM

Total GB/s Read-GB/s Write-GB/s

Internal Use - Confidential 228 Copyright © Dell Inc. All Rights Reserved.
 2 - Data Load 100GbE Total Time Data Load (hh:mm:ss)
Time to Load [Link]
[Link]

12 VMs Test [Link]

[Link] [Link]
[Link]
[Link] [Link] [Link] [Link] [Link] [Link] [Link] [Link] [Link] [Link]
[Link]
[Link]
[Link]
100GbE accounts for a [Link]
46.66% reduction in the [Link]
VM-1 VM-2 VM-3 VM-4 VM-5 VM-6 VM-7 VM-8 VM-9 VM-10 VM-11 VM-12 Average
time to Load the Data
Series1
Lower load time is better
25GbE Total Time Data Load (hh:mm:ss)
[Link] [Link] [Link] [Link] [Link] [Link] [Link] [Link] [Link] [Link]
[Link] [Link] [Link] [Link]
[Link]
[Link]
[Link]
[Link]
[Link]
[Link]
[Link]
VM-1 VM-2 VM-3 VM-4 VM-5 VM-6 VM-7 VM-8 VM-9 VM-10 VM-11 VM-12 Average

Series1

Internal Use - Confidential 229 Copyright © Dell Inc. All Rights Reserved.
 2 – Data Load • This is where we see our largest improvement as the load
Conclusion
process is throughput intensive
• 100GbE is nearly 2x as fast as 25GbE ; there is a
reduction of 46.66% in the time required to load
• There is greater throughput with 100GbE 18.94 vs 12.88
GB\s in 25GbE
• Similar workloads where data is being loaded/extracted
to/from DBs can see similar improvements e.g database
level backups

Internal Use - Confidential 230 Copyright © Dell Inc. All Rights Reserved.
 3 – Data Query • TPC-DS contains a set of 99 queries with wide variation in complexity
Overview and range of data scanned. Each TPC-DS query asks a business
question and includes the corresponding query to answer the question
– time to complete is the measurement
• Of the 99 we focused on 77 as we had full set of results for these 77.
This is the third step in
the workflow • Test can be run multi-threaded in our case we ran 16 threads to comply
with previous tests
• 4 sets of test on 25G and 100G
• 1 VM running queries with no other vms powered on
• 6 VMs (1 per Host) running queries
• 12 VMs (2 per Host) running queries
• 18 VMs (3 per Host) running queries

Internal Use - Confidential 231 Copyright © Dell Inc. All Rights Reserved.
 3 – Data Query
Overview
cont’d

• What would we measure?

1. Throughput while querying
2. Average and Total time to run the 77 Queries

Internal Use - Confidential 232 Copyright © Dell Inc. All Rights Reserved.
 3 – Data Query
100G VSAN Throughput Querying 1
Single VM Max Throughput 2.23 GB\s
Throughput 2.5
2
1.5

GB/s
1
0.5
0
Test #1 [Link] AM[Link] PM[Link] PM[Link] PM[Link] PM[Link] AM[Link] AM[Link] AM[Link] AM

Throughput is relatively Total Throughput Read-GB/s Write-GB/s

low and similar
25G VSAN Throughput Querying 1Max Throughput 2.16 GB\s
2.5
2
1.5
GB/s

1
0.5
0
[Link] AM[Link] AM[Link] AM[Link] PM[Link] PM[Link] PM[Link] PM[Link] AM[Link] AM

Total Throughput Read-GB/s Write-GB/s

Internal Use - Confidential 233 Copyright © Dell Inc. All Rights Reserved.
 3 – Data Average Query Time Single Total Query Time Single VM
Query Single VM (minutes) Single VM (minutes)
VM Timings 60.00
4500.00
52.04 4007.31
4000.00
50.00
3500.00
40.80 3143.00
40.00 3000.00
Test #1
2500.00
100GbE is 24% quicker 30.00
to run all queries 2000.00

20.00 1500.00
Lower time equates to
faster and better 1000.00
10.00
500.00

0.00 0.00
25G 100G 25G 100G

Average Query Time Total Query Time Single VM

4007.31 = 66 Hours
3143.00 = 52 Hours

Internal Use - Confidential 234 Copyright © Dell Inc. All Rights Reserved.
 3 – Data 100GbE vSAN Throughput Querying 18
Query 18 VMs 20 Max Throughput 17.83 GB/s

Throughput 15

GB/s
10

Test #4 0
[Link] PM[Link] PM[Link] AM[Link] PM[Link] PM[Link] AM[Link] AM[Link] PM[Link] PM
There is 24.25% more
throughput on the Total Throughput Read-GB/s Write-GB/s
100GbE System 25GbE vSAN Throughput Querying 18
16 Max Throughput 14.35 GB/s
14
12
10
GB/s

8
6
4
2
0
[Link] AM[Link] AM[Link] PM[Link] PM[Link] AM[Link] AM[Link] PM[Link] PM[Link] PM

Total Throughput Read-GB/s Write-GB/s

Internal Use - Confidential 235 Copyright © Dell Inc. All Rights Reserved.
 3 – Data Average Query Time 18 VM Total Query Time Single VM
(minutes) 18 VM (minutes)
Query 18 VMs
90.00 lower is better 7000.00
lower is better
Timings 80.00 77.74
6000.00
6057.71

69.30 5385.85
70.00
5000.00
60.00
Test #4
50.00 4000.00
100G is 11.10% quicker
to run all queries – that’s 40.00 3000.00
11.10% for each of the
30.00
18 VM’s 2000.00
20.00
1000.00
10.00

0.00 0.00
25G 100G 25G 100G

Average Query Time Total Query Time Single VM

6057.71 = 100.96 Hours

5385.85 = 89.76 Hours
Internal Use - Confidential 236 Copyright © Dell Inc. All Rights Reserved.
 Average Query Time (Minutes)
3 – Data 90.00
Query 80.00
70.00 63.00
77.74
69.30

60.00 52.04 55.00 54.06

47.89
50.00 40.80
40.00
30.00
20.00
Other Query Results 10.00
0.00
Summary 1VM 6 VMs 12 VMS 18 VMs

25G 100G
Total Query Time (Minutes)
7000.00
6057.71
6000.00 5385.85
4923.00
5000.00 4225.08
4007.31 4203.00
4000.00 3671.00
3143.00
3000.00
2000.00
1000.00
0.00
1VM 6 VMs 12 VMs 18 VMs
24% Quicker 12.66% Quicker 14.18% Quicker 11.10% Quicker

25G 100G

Internal Use - Confidential 237 Copyright © Dell Inc. All Rights Reserved.
 3 – Data Main Savings
Query Main [Link]
areas of
saving [Link] [Link]

[Link] [Link]

Data Load sees a [Link]

decrease in time
required by 46.66% [Link]

[Link] [Link]
Data Query sees a
decrease in time [Link]
[Link]
required by 14.18%
[Link]

[Link]
DataLoad DataQuery

25G 100G

Internal Use - Confidential 238 Copyright © Dell Inc. All Rights Reserved.