NFNN2, 20th-21st June 2005

National e-Science Centre, Edinburgh

TCP Tuning Techniques for

High-Speed Wide-Area Networks

Brian L. Tierney

Distributed Systems Department

Lawrence Berkeley National Laboratory

http://gridmon.dl.ac.uk/nfnn/

Wizard Gap

Slide from
Matt Mathis, PSC
Brian L. Tierney Slide: 2

1
 Today’s Talk

 This talk will cover:

 Information needed to be a “wizard”
 Current work being done so you don’t have to be a
wizard

 Outline
 TCP Overview
 TCP Tuning Techniques (focus on Linux)
 TCP Issues
 Network Monitoring Tools
 Current TCP Research

Brian L. Tierney Slide: 3

How TCP works: A very short overview

 Congestion window (CWND) = the number of packets the

sender is allowed to send
 The larger the window size, the higher the throughput
 Throughput = Window size / Round-trip Time
 TCP Slow start
 exponentially increase the congestion window size until a packet is
lost
 this gets a rough estimate of the optimal congestion window
size
packet loss
timeout
CWND
slow start:
exponential congestion retransmit:
increase avoidance: slow start
linear again
increase

time
Brian L. Tierney Slide: 4

2
 TCP Overview

 Congestion avoidance
 additive increase: starting from the rough estimate, linearly
increase the congestion window size to probe for additional
available bandwidth
 multiplicative decrease: cut congestion window size aggressively if
a timeout occurs

packet loss
timeout
CWND
slow start:
exponential congestion retransmit:
increase avoidance: slow start
linear again
increase

time

Brian L. Tierney Slide: 5

TCP Overview
 Fast Retransmit: retransmit after 3 duplicate acks (got 3 additional
packets without getting the one you are waiting for)
 this prevents expensive timeouts
 no need to go into “slow start” again
 At steady state, CWND oscillates around the optimal window size
 With a retransmission timeout, slow start is triggered again

packet loss
timeout
CWND
slow start:
exponential congestion retransmit:
increase avoidance: slow start
linear again
increase

time

Brian L. Tierney Slide: 6

3
 Terminology

 The term “Network Throughput” is vague and should be

avoided
 Capacity: link speed
 Narrow Link: link with the lowest capacity along a path
 Capacity of the end-to-end path = capacity of the narrow link
 Utilized bandwidth: current traffic load
 Available bandwidth: capacity – utilized bandwidth
 Tight Link: link with the least available bandwidth in a path
 Achievable bandwidth: includes protocol and host issues

45 Mbps 10 Mbps 100 Mbps 45 Mbps

source sink

Narrow Link
Tight Link
Brian L. Tierney Slide: 7

More Terminology

 RTT: Round-trip time

 Bandwidth*Delay Product = BDP
 The number of bytes in flight to fill the entire path
 Example: 100 Mbps path; ping shows a 75 ms RTT
BDP = 100 * 0.075 / 2 = 3.75 Mbits (470 KB)
 LFN: Long Fat Networks
 A network with a large BDP

Brian L. Tierney Slide: 8

4
 TCP Performance Tuning Issues

 Getting good TCP performance over high-latency

high-bandwidth networks is not easy!
 You must keep the pipe full, and the size of the
pipe is directly related to the network latency
 Example: from LBNL (Berkeley, CA) to ANL (near
Chicago, IL), the narrow link is 1000 Mbits/sec, and the
one-way latency is 25ms
 Need (1000 / 8) * .025 sec = 3.125 MBytes of data “in
flight” to fill the pipe

Brian L. Tierney Slide: 9

Setting the TCP buffer sizes

 It is critical to use the optimal TCP send and receive

socket buffer sizes for the link you are using.
 Recommended size = 2 x Bandwidth Delay Product (BDP)
 if too small, the TCP window will never fully open up
 if too large, the sender can overrun the receiver, and the TCP
window will shut down
 Default TCP buffer sizes are way too small for this type of
network
 default TCP send/receive buffers are typically 64 KB
 with default TCP buffers, you can only get a small % of the
available bandwidth!

Brian L. Tierney Slide: 10

5
 Importance of TCP Tuning
Throughput (Mbits/sec)
Tuned for Tuned for Tuned for
300 LAN WAN Both
264 264

200
152
112 112
100
44

64KB TCP 512 KB

Buffers TCP Buffers

LAN (rtt = 1ms)

WAN (rtt = 50ms)
Brian L. Tierney Slide: 11

TCP Buffer Tuning: System

 Need to adjust system max TCP buffer
 Example: in Linux (2.4 and 2.6) add the entries below to the file
/etc/sysctl.conf, and then run "sysctl -p”

# increase TCP max buffer size

net.core.rmem_max = 16777216
net.core.wmem_max = 16777216
# increase Linux autotuning TCP buffer limits
# min, default, and max number of bytes to use
net.ipv4.tcp_rmem = 4096 87380 16777216
net.ipv4.tcp_wmem = 4096 65536 16777216

 Similar changes needed for other Unix OS’s

 For more info, see: http://dsd.lbl.gov/TCP-Tuning/

Brian L. Tierney Slide: 12

6
 TCP Buffer Tuning: Application
 Must adjust buffer size in your applications:
int skt, int sndsize = 2 * 1024 * 1024;
err = setsockopt(skt, SOL_SOCKET, SO_SNDBUF,
(char *)&sndsize,(int)sizeof(sndsize));
and/or
err = setsockopt(skt, SOL_SOCKET, SO_RCVBUF,
(char *)&sndsize,(int)sizeof(sndsize));

 It’s a good idea to check the following:

err = getsockopt(skt, SOL_SOCKET, SO_RCVBUF,
(char *)&sockbufsize, &size);
If (size != sndsize)
printf(stderr, “Warning: requested TCP buffer of %d,
but only got %d \n”, sndsize, size);

Brian L. Tierney Slide: 13

Determining the Buffer Size

 The optimal buffer size is twice the bandwidth*delay

product of the link:
buffer size = 2 * bandwidth * delay
 The ping program can be used to get the delay
 e.g.: >ping -s 1500 lxplus.cern.ch
1500 bytes from lxplus012.cern.ch: icmp_seq=0. time=175. ms
1500 bytes from lxplus012.cern.ch: icmp_seq=1. time=176. ms
1500 bytes from lxplus012.cern.ch: icmp_seq=2. time=175. ms

 pipechar or pathrate can be used to get the bandwidth of

the slowest hop in your path. (see next slides)
 Since ping gives the round trip time (RTT), this formula
can be used instead of the previous one:
buffer size = bandwidth * RTT

Brian L. Tierney Slide: 14

7
 Buffer Size Example
 ping time = 50 ms
 Narrow link = 500 Mbps (62 Mbytes/sec)
 e.g.: the end-to-end network consists of all 1000 BT
ethernet and OC-12 (622 Mbps)
 TCP buffers should be:
 .05 sec * 62 = 3.1 Mbytes

Brian L. Tierney Slide: 15

Sample Buffer Sizes

 UK to...
 UK (RTT = 5 ms, narrow link = 1000 Mbps) : 625 KB
 Europe: (RTT = 25 ms, narrow link = 500 Mbps): 1.56 MB
 US: (RTT = 150 ms, narrow link = 500 Mbps): 9.4 MB
 Japan: (RTT = 260, narrow link = 150 Mbps): 4.9 MB

 Note: default buffer size is usually only 64 KB, and default

maximum buffer size for is only 256KB
 Linux Autotuning default max = 128 KB;
 10-150 times too small!

 Home DSL, UK to US (RTT = 150, narrow link = 1 Mbps): 19 KB

 Default buffers are OK.

Brian L. Tierney Slide: 16

8
 More Problems: TCP congestion control

Path =
LBL to
CERN
(Geneva)
OC-3, (in
2000), RTT
= 150 ms

average
BW =
30 Mbps

Brian L. Tierney Slide: 17

Work-around: Use Parallel Streams

RTT = 70 ms

graph from Tom Dunigan, ORNL

Brian L. Tierney Slide: 18

9
 Tuned Buffers vs. Parallel Steams

Throughput (Mbits/sec)
30
25
20
15
10
5
0
no tuning tuned 10 tuned
TCP parallel TCP
buffers streams, buffers, 3
no tuning parallel
streams

Brian L. Tierney Slide: 19

Parallel Streams Issues

 Potentially unfair
 Places more load
on the end hosts
 But they are
necessary when
you don’t have root
access, and can’t
convince the
sysadmin to
increase the max
TCP buffers
graph from Tom Dunigan, ORNL

Brian L. Tierney Slide: 20

10
 NFNN2, 20th-21st June 2005
National e-Science Centre, Edinburgh

Network Monitoring Tools

http://gridmon.dl.ac.uk/nfnn/

traceroute
>traceroute pcgiga.cern.ch
traceroute to pcgiga.cern.ch (192.91.245.29), 30 hops max, 40 byte packets
1 ir100gw-r2.lbl.gov (131.243.2.1) 0.49 ms 0.26 ms 0.23 ms
2 er100gw.lbl.gov (131.243.128.5) 0.68 ms 0.54 ms 0.54 ms
3 198.129.224.5 (198.129.224.5) 1.00 ms *d9* 1.29 ms
4 lbl2-ge-lbnl.es.net (198.129.224.2) 0.47 ms 0.59 ms 0.53 ms
5 snv-lbl-oc48.es.net (134.55.209.5) 57.88 ms 56.62 ms 61.33 ms
6 chi-s-snv.es.net (134.55.205.102) 50.57 ms 49.96 ms 49.84 ms
7 ar1-chicago-esnet.cern.ch (198.124.216.73) 50.74 ms 51.15 ms 50.96
ms
8 cernh9-pos100.cern.ch (192.65.184.34) 175.63 ms 176.05 ms 176.05
ms
9 cernh4.cern.ch (192.65.185.4) 175.92 ms 175.72 ms 176.09 ms
10 pcgiga.cern.ch (192.91.245.29) 175.58 ms 175.44 ms 175.96 ms

Can often learn about the network from the router names:
ge = Gigabit Ethernet
oc48 = 2.4 Gbps (oc3 = 155 Mbps, oc12=622 Mbps)

Brian L. Tierney Slide: 22

11
 Iperf
 iperf : very nice tool for measuring end-to-end TCP/UDP performance
 http://dast.nlanr.net/Projects/Iperf/
 Can be quite intrusive to the network
 Example:
 Server: iperf -s -w 2M
 Client: iperf -c hostname -i 2 -t 20 -l 128K -w 2M
Client connecting to hostname
[ ID] Interval Transfer Bandwidth
[ 3] 0.0- 2.0 sec 66.0 MBytes 275 Mbits/sec
[ 3] 2.0- 4.0 sec 107 MBytes 451 Mbits/sec
[ 3] 4.0- 6.0 sec 106 MBytes 446 Mbits/sec
[ 3] 6.0- 8.0 sec 107 MBytes 443 Mbits/sec
[ 3] 8.0-10.0 sec 106 MBytes 447 Mbits/sec
[ 3] 10.0-12.0 sec 106 MBytes 446 Mbits/sec
[ 3] 12.0-14.0 sec 107 MBytes 450 Mbits/sec
[ 3] 14.0-16.0 sec 106 MBytes 445 Mbits/sec
[ 3] 16.0-24.3 sec 58.8 MBytes 59.1 Mbits/sec
[ 3] 0.0-24.6 sec 871 MBytes 297 Mbits/sec

Brian L. Tierney Slide: 23

pathrate / pathload

 Nice tools from Georgia Tech:

 pathrate: measures the capacity of the narrow link
 pathload: measures the available bandwidth

 Both work pretty well.

 pathrate can take a long time (up to 20 minutes)
 These tools attempt to be non-intrusive
 Open Source; available from:
 http://www.pathrate.org/

Brian L. Tierney Slide: 24

12
 pipechar

 Tool to measure hop-by-hop available bandwidth,

capacity, and congestion
 Takes 1-2 minutes to measure an 8 hop path
 client-side only tool: puts very little load on the network
(about 100 Kbits/sec)
 But not always accurate
 Results affected by host speed
 Hard to measure links faster than host interface
 Results after a slow hop typically not accurate, for example, if
the first hop is a wireless link, and all other hops are 100 BT or
faster, then results are not accurate
 Available from: http://dsd.lbl.gov/NCS/
 part of the netest package

Brian L. Tierney Slide: 25

pipechar output

dpsslx04.lbl.gov(59)>pipechar firebird.ccs.ornl.gov
PipeChar statistics: 82.61% reliable
From localhost: 827.586 Mbps GigE (1020.4638 Mbps)
1: ir100gw-r2.lbl.gov (131.243.2.1 )
| 1038.492 Mbps GigE <11.2000% BW used>
2: er100gw.lbl.gov (131.243.128.5)
| 1039.246 Mbps GigE <11.2000% BW used>
3: lbl2-ge-lbnl.es.net (198.129.224.2)
| 285.646 Mbps congested bottleneck <71.2000% BW used>
4: snv-lbl-oc48.es.net (134.55.209.5)
| 9935.817 Mbps OC192 <94.0002% BW used>
5: orn-s-snv.es.net (134.55.205.121)
| 341.998 Mbps congested bottleneck <65.2175% BW used>
6: ornl-orn.es.net (134.55.208.62)
| 298.089 Mbps congested bottleneck <70.0007% BW used>
7: orgwy-ext.ornl.gov (192.31.96.225)
| 339.623 Mbps congested bottleneck <65.5502% BW used>
8: ornlgwy-ext.ens.ornl.gov (198.124.42.162)
| 232.005 Mbps congested bottleneck <76.6233% BW used>
9: ccsrtr.ccs.ornl.gov (160.91.0.66 )
| 268.651 Mbps GigE (1023.4655 Mbps)
10: firebird.ccs.ornl.gov (160.91.192.165)
Brian L. Tierney Slide: 26

13
 tcpdump / tcptrace

 tcpdump: dump all TCP header information for a specified

source/destination
 ftp://ftp.ee.lbl.gov/

 tcptrace: format tcpdump output for analysis using xplot

 http://www.tcptrace.org/
 NLANR TCP Testrig : Nice wrapper for tcpdump and tcptrace tools
 http://www.ncne.nlanr.net/TCP/testrig/

 Sample use:
tcpdump -s 100 -w /tmp/tcpdump.out host hostname
tcptrace -Sl /tmp/tcpdump.out
xplot /tmp/a2b_tsg.xpl

Brian L. Tierney Slide: 27

tcptrace and xplot

 X axis is time
 Y axis is sequence number
 the slope of this curve gives the throughput over
time.
 xplot tool make it easy to zoom in

Brian L. Tierney Slide: 28

14
 Zoomed In View
 Green Line: ACK values received from the receiver
 Yellow Line tracks the receive window advertised from the receiver
 Green Ticks track the duplicate ACKs received.
 Yellow Ticks track the window advertisements that were the same as
the last advertisement.
 White Arrows represent segments sent.
 Red Arrows (R) represent retransmitted segments

Brian L. Tierney Slide: 29

Other Tools

 NLANR Tools Repository:

 http://www.ncne.nlanr.net/software/tools/

 SLAC Network MonitoringTools List:

 http://www.slac.stanford.edu/xorg/nmtf/nmtf-tools.html

Brian L. Tierney Slide: 30

15
 Other TCP Issues

 Things to be aware of:

 TCP slow-start
On a path with a 50 ms RTT, it takes 12 RTT’s to
ramp up to full window size, so need to send about
10 MB of data before the TCP congestion window
will fully open up.

 host issues
Memory copy speed
I/O Bus speed
Disk speed

Brian L. Tierney Slide: 31

TCP Slow Start

Brian L. Tierney Slide: 32

16
 Duplex Mismatch Issues
 A common source of trouble with Ethernet
networks is that the host is set to full duplex, but
the Ethernet switch is set to half-duplex, or visa
versa.
 Most newer hardware will auto-negotiate this, but
with some older hardware, auto-negotiation
sometimes fails
 result is a working but very slow network (typically only
1-2 Mbps)
 best for both to be in full duplex if possible, but some
older 100BT equipment only supports half-duplex
 NDT is a good tool for finding duplex issues:
 http://e2epi.internet2.edu/ndt/

Brian L. Tierney Slide: 33

Jumbo Frames

 Standard Ethernet packet is 1500 bytes (aka: MTU)

 Some gigabit Ethernet hardware supports “jumbo frames”
(jumbo packet) up to 9 KBytes
 This helps performance by reducing the number of host interrupts
 Some jumbo frame implementations do not interoperate
 Most routers allow at most 4K MTUs

 First Ethernet was 3 Mbps (1972)

 First 10 Gbit/sec Ethernet hardware: 2001
 Ethernet speeds have increased 3000x since the 1500 byte frame
was defined
 Computers now have to work 3000x harder to keep the network full

Brian L. Tierney Slide: 34

17
 Linux Autotuning

 Sender-side TCP buffer autotuning introduced in Linux 2.4

 TCP send buffer starts at 64 KB
 As the data transfer takes place, the buffer size is continuously re-
adjusted up max autotune size (default = 128K)
 Need to increase defaults: (in /etc/sysctl.conf)
# increase TCP max buffer size
net.core.rmem_max = 16777216
net.core.wmem_max = 16777216
# increase Linux autotuning TCP buffer limits
# min, default, and max number of bytes to use
net.ipv4.tcp_rmem = 4096 87380 16777216
net.ipv4.tcp_wmem = 4096 65536 16777216

 Receive buffers need to be bigger than largest send buffer used

 Use setsockopt() call

Brian L. Tierney Slide: 35

Linux 2.4 Issues

 ssthresh caching
 ssthresh (Slow Start Threshold): size of CWND to use when
switching from exponential increase to linear increase
 The value for ssthresh for a given path is cached in the routing
table.
 If there is a retransmission on a connection to a given host, then
all connections to that host for the next 10 minutes will use a
reduced ssthresh.
 Or, if the previous connect to that host is particularly good, then
you might stay in slow start longer, so it depends on the path
 The only way to disable this behavior is to do the following before
all new connections (you must be root):
 sysctl -w net.ipv4.route.flush=1
 The web100 kernel patch adds a mechanism to permanently
disable this behavior:
 sysctl -w net.ipv4.web100_no_metrics_save = 1

Brian L. Tierney Slide: 36

18
 ssthresh caching

•The value of
CWND when
this loss
happened will
get cached

Brian L. Tierney Slide: 37

Linux 2.4 Issues (cont.)

 SACK implementation problem
 For very large BDP paths where the TCP window is > 20 MB, you
are likely to hit the Linux SACK implementation problem.
 If Linux has too many packets in flight when it gets a SACK event,
it takes too long to located the SACKed packet,
 you get a TCP timeout and CWND goes back to 1 packet.
 Restricting the TCP buffer size to about 12 MB seems to avoid this
problem, but limits your throughput.
 Another solution is to disable SACK.
sysctl -w net.ipv4.tcp_sack = 0
 This is still a problem in 2.6, but they are working on a solution
 Transmit queue overflow
 If the interface transmit queue overflows, the Linux TCP stack
treats this as a retransmission.
 Increasing txqueuelen can help:
ifconfig eth0 txqueuelen 1000

Brian L. Tierney Slide: 38

19
 NFNN2, 20th-21st June 2005
National e-Science Centre, Edinburgh

Recent/Current TCP Work

http://gridmon.dl.ac.uk/nfnn/

TCP Response Function

 Well known fact that TCP does not scale to high-

speed networks
 Average TCP congestion window = 1.2 p
segments
 p = packet loss rate
 What this means:
 For a TCP connection with 1500-byte packets and a
100 ms round-trip time, filling a 10 Gbps pipe would
require a congestion window of 83,333 packets, and a
packet drop rate of at most one drop every
5,000,000,000 packets.
 requires at most one packet loss every 6000s, or
1h:40m to keep the pipe full

Brian L. Tierney Slide: 40

20
 Proposed TCP Modifications

 High Speed TCP: Sally Floyd

 http://www.icir.org/floyd/hstcp.html
 BIC/CUBIC:
 http://www.csc.ncsu.edu/faculty/rhee/export/bitcp/
 LTCP (Layered TCP)
 http://students.cs.tamu.edu/sumitha/research.html
 HTCP: (Hamilton TCP)
 http://www.hamilton.ie/net/htcp/
 Scalable TCP
 http://www-lce.eng.cam.ac.uk/~ctk21/scalable/

Brian L. Tierney Slide: 41

Proposed TCP Modifications (cont.)

 XCP:
 XCP rapidly converges on the optimal congestion window using a
completely new router paradigm.
 This makes it very difficult to deploy and test
 http://www.ana.lcs.mit.edu/dina/XCP/
 FAST TCP:
 http://netlab.caltech.edu/FAST/

 Each of these alternatives give roughly similar throughput

 Vary mainly in “stability” and “friendliness” with other protocols
 Each of these require sender-side only modifications to
standard TCP

Brian L. Tierney Slide: 42

21
 TCP: Reno vs. BIC

 TCP-Reno
(Linux 2.4)

 BIC-TCP
(Linux 2.6)

Brian L. Tierney Slide: 43

TCP: Reno vs. BIC

• BIC-TCP
recovers from
loss more
aggressively
than TCP-Reno

Brian L. Tierney Slide: 44

22
 Sample Results
From Doug Leith, Hamilton Institute, http://www.hamilton.ie/net/eval/

Fairness
Between Flows

Link Utilization

Brian L. Tierney Slide: 45

New Linux 2.6 changes

 Added receive buffer autotuning: adjust receive window
based on RTT
 sysctl net.ipv4.tcp_moderate_rcvbuf
 Still need to increase max value: net.ipv4.tcp_rmem
 Starting in Linux 2.6.7 (and back-ported to 2.4.27), BIC
TCP is part of the kernel, and enabled by default.
 Bug found that caused performance problems under some
circumstances, fixed in 2.6.11.
 Added ability to disable ssthresh caching (like web100)
net.ipv4.tcp_no_metrics_save = 1

Brian L. Tierney Slide: 46

23
 Linux 2.6 Issues

 "tcp segmentation offload” issue:

 Linux 2.6 ( < 2.6.11) has bug with certain Gigabit and
10 Gig ethernet drivers and NICs that support "tcp
segmentation offload",
 These include Intel e1000 and ixgb drivers,
Broadcom tg3, and the s2io 10 GigE drivers.
To fix this problem, use ethtool to disable
segmentation offload:
ethtool -K eth0 tso off
 Bug fixed in Linux 2.6.12

Brian L. Tierney Slide: 47

Linux 2.6.12 Results

• BIC-TCP is ON by default in Linux 2.6
• un-tuned results up to 100x faster!
Path Linux 2.4 Linux 2.4 Linux 2.6 Linux 2.6,
Un-tuned Hand-tuned with BIC no BIC
LBL to ORNL 10 Mbps 300 Mbps 700 Mbps 500 Mbps
RTT = 67 ms
LBL to PSC 8 Mbps 300 Mbps 830 Mbps 625 Mbps
RTT = 83 ms

LBL to IE 4 Mbps 70 Mbps 560 Mbps 140 Mbps

RTT = 153 ms

• Results = Peak Speed during 3 minute test

• Must increase default max TCP send/receive buffers
• Sending host = 2.8 GHz Intel Xeon with Intel e1000 NIC

Brian L. Tierney Slide: 48

24
 Linux 2.6.12-rc3 Results

Note that BIC reaches

Max throughput MUCH
faster

RTT = 67 ms

Brian L. Tierney Slide: 49

Remaining Linux BIC Issues

 But: on
some
paths BIC
still seems
to have
problems…

RTT = 83 ms

Brian L. Tierney Slide: 50

25
 NFNN2, 20th-21st June 2005
National e-Science Centre, Edinburgh

Application Performance Issues

http://gridmon.dl.ac.uk/nfnn/

Techniques to Achieve High

Throughput over a WAN
 Consider using multiple TCP sockets for the data
stream
 Use a separate thread for each socket
 Keep the data pipeline full
 use asynchronous I/O
overlap I/O and computation
 read and write large amounts of data (> 1MB) at a time
whenever possible
 pre-fetch data whenever possible
 Avoid unnecessary data copies
 manipulate pointers to data blocks instead
Brian L. Tierney Slide: 52

26
 Use Asynchronous I/O
 I/O followed by
processing

Next IO starts
when processing
ends
 overlapped I/O and
processing

process
previous block

almost a 2:1 speedup remote IO

Brian L. Tierney Slide: 53

Throughput vs. Latency

 Most of the techniques we have discussed are designed

to improve throughput
 Some of these might increase latency
 with large TCP buffers, OS will buffer more data before sending it
 Goal of a distributed application programmer:
 hide latency
 Some techniques to help decrease latency:
 use separate control and data sockets
 use TCP_NODELAY option on control socket
 combine control messages together into 1 larger message
whenever possible on TCP_NODELAY sockets

Brian L. Tierney Slide: 54

27
 scp Issues

 Don’t use scp to copy large files!

 scp has its own internal buffering/windowing that
prevents it from ever being able to fill LFNs!
 Explanation of problem and openssh patch
solution from PSC
 http://www.psc.edu/networking/projects/hpn-ssh/

Brian L. Tierney Slide: 55

Conclusions
 The wizard gap is starting to close (slowly)
 If max TCP buffers are increased
 Tuning TCP is not easy!
 no single solution fits all situations
 need to be careful TCP buffers are not too big or too small
 sometimes parallel streams help throughput, sometimes they hurt
 Linux 2.6 helps a lot
 Design your network application to be as flexible as possible
 make it easy for clients/users to set the TCP buffer sizes
 make it possible to turn on/off parallel socket transfers
 probably off by default
 Design your application for the future
 even if your current WAN connection is only 45 Mbps (or less), some day
it will be much higher, and these issues will become even more important

Brian L. Tierney Slide: 56

28
 For More Information

http://dsd.lbl.gov/TCP-tuning/
 links to all network tools mentioned here
 sample TCP buffer tuning code, etc.

[email protected]

Brian L. Tierney Slide: 57

29