MPI-1

Lecture 3
January 15, 2024
Adapted from Neha Karanjkar’s slides

2
Parallel Execution – Single Node

Core
Process
Memory

$ ps
Instruction 1 Instruction 1 Instruction 1 Instruction 1
Instruction 2 Instruction 2 Instruction 2 Instruction 2 parallel.x
… … … … parallel.x
parallel.x parallel.x parallel.x parallel.x parallel.x
parallel.x

mpirun -np 4 ./parallel.x

3
 Access Latency (Intel KNL)

Simulating the Network Activity of

Modern Manycores, IEEE Access 2019 4
 Cores

Simulating the Network Activity of

Modern Manycores, IEEE Access 2019 5
Parallel Execution – Multiple
Nodes

Core
Process
Memory

Instruction 1 Instruction 1 Instruction 1 Instruction 1

Instruction 2 Instruction 2 Instruction 2 Instruction 2
… … … …

parallel.x parallel.x parallel.x parallel.x SIMD

parallel.x parallel.x parallel.x parallel.x

mpirun -np 8 ./parallel.x

6
MPI Code On Multiple Nodes
• mpirun -np 8 –f hostfile ./program.x
• Round robin process placement
• mpirun -np 8 –hosts h1,h2,h3,h4 ./program.x
• Host-wise placement
• mpirun -np 8 –hosts h1:2,h2:2,h3:2,h4:2 ./program.x

7
Parallel Execution – Multiple
Nodes

Core
Process
Memory

Instruction 1 Instruction 1 Instruction 1 Instruction 1

Instruction 2 Instruction 2 Instruction 2 Instruction 2
… … … …

parallel.x parallel.x parallel.x parallel.x

parallel.x parallel.x parallel.x parallel.x

mpirun -np 8 -hosts h1,h2,h3,h4 ./program.x

mpirun -np 8 -hosts h1:2,h2:2,h3:2,h4:2 ./program.x
8
Parallel Sum (Optimized)

Source: GGKK Chapter 5 9

Getting Started

• gather information about the

parallel job
• set up internal library state
• prepare for communication

Initialization

Finalization

10
 MPI Process Identification
Global
communicator

Total number
of processes

Rank of a
process

11
 Entities
Communicator (communication handle)
• Defines the scope
• Specifies communication context

Process
• Belongs to a group
• Identified by a rank within a group

Identification
• MPI_Comm_size – total number of processes in communicator
• MPI_Comm_rank – rank in the communicator
12
MPI_COMM_WORLD

• Required in every MPI communication

• Process identified by rank/id

3
2

0 4

13
 Communicator
• Communication handle among a group/collection of processes
• Representative of communication domain
• Associated with a context ID (in MPICH)
• Predefined:
• MPI_COMM_WORLD
• MPI_COMM_SELF

14
Output
// get number of tasks
MPI_Comm_size (MPI_COMM_WORLD, &numtasks);

// get my rank
MPI_Comm_rank (MPI_COMM_WORLD, &rank);

printf (“Hello I’m rank %d of %d processes\n”, rank, numtasks);

mpiexec –n 4 –hosts host1,host2,host3,host4 ./exe

mpiexec –n 4 –hosts host1:2,host2:2,host3:2,host4:2 ./exe
mpiexec –n 8 –hosts host1,host2,host3,host4 ./exe
mpiexec –n 8 –hosts host1:1,host2:2,host3:3,host4:4 ./exe
15
Host

16
Process Launch
Memory

MPI MPI MPI MPI

process process process process
Launch node

Reading (optional):
A Scalable Process-Management Environment for Parallel Programs
17
Parallel Environment
• Job scheduler
• Resources to run the parallel job
• Process manager
• Starts and stops the processes
• Parallel library
• Application calls for communication

18
Multiprocess Daemon

Butler et al., EuroPVM/MPI2000

19
 Process Management Setup
Parallel program
library (e.g. MPI)
Process
management
interface (PMI)

Resource manager/
Job scheduler/
Process Manager

Reading (optional):
PMI: A Scalable Parallel Process-Management Interface for Extreme-Scale Systems 20
Process Manager

• Start and stop processes in a scalable way

• Setup communication channels for parallel processes
• stdin/stdout/stderr forwarding
• ./a.out 2>efile

21
Process Management Interface

• Interface between process manager and MPI library

• Processes can exchange information about peers by
querying PMI
• Uses key-value store for process-related data

22
Process Launch
Memory

MPI MPI MPI MPI

process process process process

PMI PMI PMI PMI

23
Process Launch
Memory

PMI PMI PMI PMI

KVS $ KVS $ KVS $ KVS $

24
 Hydra Process Manager
• A process management system for starting parallel jobs
• Uses existing daemons (viz. ssh) to start MPI processes
• Automatically detects resource managers if any and interacts
with them
• $ mpiexec ./app
• Hydra gets information about allocated resources and
launches processes
• Passes environment variables from the shell on which
mpiexec is launched to the launched processes

There are others – gforker, slurm, etc.

25
mpiexec
Memory

host1 host2 host3 host4

Launch node
Proxy Proxy Proxy Proxy

mpiexec –n 4 –hosts host1,host2,host3,host4 ./exe

26
 Launch Node
mpiexec -np 8 -hosts host1:3,host2:3,host3:3 ./exe

27
Compute Node Processes

28
Hydra and mpiexec

Source: wiki.mpich.org
29
MPI Process Management
Initializes
and queries
PMI

30
csews*
lscpu

31
 Multiple Tasks Core ID

Get the
core ID

32
Process Placement

33
Process Placement

34
 Executing MPI programs
• Check `which mpicc` and `which mpiexec`
• Update PATH environment variable
• Compile
• mpicc –o filename filename.c
• Run
• mpiexec –np 4 –f hostfile filename [Often “No such file” error]
• mpiexec –np 4 –f hostfile ./filename

35
Resources for MPI

• Marc Snir, Steve W. Otto, Steven Huss-Lederman, David W.

Walker and Jack Dongarra,
MPI - The Complete Reference, Second Edition, Volume 1, The
MPI Core.
• William Gropp, Ewing Lusk, Anthony Skjellum,
Using MPI: portable parallel programming with the message-
passing interface, 3rd Ed., Cambridge MIT Press, 2014.
• https://www.mpi-forum.org/docs/mpi-4.0/mpi40-report.pdf

36
Message Passing Paradigm
• Message sends and receives
• Explicit communication

Communication patterns
• Point-to-point
• Collective

37
Communication – Message Passing

Process 0 Process 1

38
Message Passing

Process 0 Process 1 Process 2 Process 3

39
MPI Basics
• Process is identified by rank
• Communicator specifies communication context

Message Envelope

Source
Destination
Communicator
Tag (0:MPI_TAG_UB)

40
MPI Data Types
• MPI_BYTE
• MPI_CHAR
• MPI_INT
• MPI_FLOAT
• MPI_DOUBLE

41
Point-to-point Communication
MPI_Send Blocking send and receive
int MPI_Send (const void *buf, int count,
MPI_Datatype datatype, int dest, int tag,
MPI_Comm comm)
SENDER

Tags should match

MPI_Recv
int MPI_Recv (void *buf, int count,
MPI_Datatype datatype, int source, int tag,
MPI_Comm comm, MPI_Status *status)
RECEIVER
42