Introduction to Parallel Programming with C and MPI at MCSR Part 1 The University of Southern Mississippi April 8, 2010.

Introduction to Parallel Programming

with C and MPI at MCSR

Part 1

The University of Southern Mississippi

April 8, 2010

What is a Supercomputer?

Loosely speaking, it is a “large” computer with an architecture that has been optimized for bigger solving problems faster than a conventional desktop, mainframe, or server computer.

- Pipelining

- Parallelism (lots of CPUs or Computers)

Supercomputers at MCSR: mimosa

- 253 CPU Intel Linux Cluster – Pentium 4- Distributed memory – 500MB – 1GB per node- Gigabit Ethernet

Supercomputers at MCSR: redwood

- 224 CPU Memory Supercomputer- Intel Itanium 2- Shared Memory: 1GB per node

Supercomputers at MCSR: sequoia

- 46 node Linux Cluster- 8 cores (CPUs) per node = 368 cores total- 2 GB memory per core (16 GB per node)- Shared memory intra-node- Distributed memory inter-node

- Intel Xeon processors

Supercomputers at MCSR: sequoia

What is Parallel Computing?

Using more than one computer (or processor) to complete a computational problem

How May a Problem be Parallelized?

Data Decomposition

Task Decomposition

Models of Parallel Programming

• Message Passing Computing– Processes coordinate and communicate results via calls to message passing

library routines– Programmers “parallelize” algorithm and add message calls– At MCSR, this is via MPI programming with C or Fortran

• Sweetgum – Origin 2800 Supercomputer (128 CPUs)• Mimosa – Beowulf Cluster with 253 Nodes• Redwood – Altix 3700 Supercomputer (224 CPUs)

• Shared Memory Computing– Processes or threads coordinate and communicate results via shared memory

variables– Care must be taken not to modify the wrong memory areas– At MCSR, this is via OpenMP programming with C or Fortran on sweetgum

Message Passing Computing at MCSR

• Process Creation• Manager and Worker Processes• Static vs. Dynamic Work Allocation • Compilation• Models• Basics• Synchronous Message Passing• Collective Message Passing• Deadlocks• Examples

Message Passing Process Creation

• Dynamic– one process spawns other processes & gives them work

– PVM

– More flexible

– More overhead - process creation and cleanup

• Static– Total number of processes determined before execution

begins

– MPI

Message Passing Processes

• Often, one process will be the manager, and the remaining processes will be the workers

• Each process has a unique rank/identifier

• Each process runs in a separate memory space and has its own copy of variables

Message Passing Work Allocation

• Manager Process– Does initial sequential processing– Initially distributes work among the workers

• Statically or Dynamically

– Collects the intermediate results from workers– Combines into the final solution

• Worker Process– Receives work from, and returns results to, the manager– May distribute work amongst themselves

(decentralized load balancing)

Message Passing Compilation

• Compile/link programs w/ message passing libraries using regular (sequential) compilers

• Fortran MPI example:include mpif.h

• C MPI example:#include “mpi.h”

Message Passing Compilation

Message Passing Models

• SPMD – Shared Program/Multiple Data– Single version of the source code used for each process– Manager executes one portion of the program; workers

execute another; some portions executed by both– Requires one compilation per architecture type– MPI

• MPMP – Multiple Program/Multiple Data– Once source code for master; another for slave– Each must be compiled separately– PVM

Message Passing Basics

• Each process must first establish the message passing environment

• Fortran MPI example:integer ierror

call MPI_INIT (ierror)

• C MPI example:MPI_Init(&argc, &argv);

Message Passing Basics

• Each process has a rank, or id number– 0, 1, 2, … n-1, where there are n processes

• With SPMD, each process must determine its own rank by calling a library routine

• Fortran MPI Example:integer comm, rank, ierrorcall MPI_COMM_RANK(MPI_COMM_WORLD, rank,

ierror)

• C MPI ExampleMPI_Comm_rank(MPI_COMM_WORLD, &rank);

Message Passing Basics

• Each process has a rank, or id number– 0, 1, 2, … n-1, where there are n processes

• Each process may use a library call to determine how many total processes it has to play with

• Fortran MPI Example:integer comm, size, ierrorcall MPI_COMM_SIZE(MPI_COMM_WORLD, size, ierror)

• C MPI ExampleMPI_Comm_size(MPI_COMM_WORLD, &size);

Message Passing Basics

• Each process has a rank, or id number– 0, 1, 2, … n-1, where there are n processes

• Once a process knows the size, it also knows the ranks (id #’s) of those other processes, and can send or receive a message to/from any other process.

• C Example:MPI_Send(buf, count, datatype, dest, tag, comm, ierror)

------DATA---------- ---EVELOPE--- -status------MPI_Recv(buf, count, datatype, sourc,tag,comm, status,ierror)

MPI Send and Receive Arguments

• Buf starting location of data• Count number of elements• Datatype MPI_Integer, MPI_Real, MPI_Character…• Destination rank of process to whom msg being sent• Source rank of sender from whom msg being received

or MPI_ANY_SOURCE

• Tag integer chosen by program to indicate type of messageor MPI_ANY_TAG

• Communicator id’s the process team, e.g., MPI_COMM_WORLD

• Status the result of the call (such as the # data items received)

Synchronous Message Passing

• Message calls may be blocking or nonblocking

• Blocking Send– Waits to return until the message has been received by the

destination process

– This synchronizes the sender with the receiver

• Nonblocking Send– Return is immediate, without regard for whether the message has

been transferred to the receiver

– DANGER: Sender must not change the variable containing the old message before the transfer is done.

– MPI_ISend() is nonblocking

Synchronous Message Passing

• Locally Blocking Send– The message is copied from the send parameter

variable to intermediate buffer in the calling process– Returns as soon as the local copy is complete– Does not wait for receiver to transfer the message from

the buffer– Does not synchronize– The sender’s message variable may safely be reused

immediately – MPI_Send() is locally blocking

Synchronous Message Passing

• Blocking Receive– The call waits until a message matching the given tag has been

received from the specified source process.– MPI_RECV() is blocking.

• Nonblocking Receive– If this process has a qualifying message waiting, retrieves that

message and returns– If no messages have been received yet, returns anyway– Used if the receiver has other work it can be doing while it waits– Status tells the receive whether the message was received– MPI_Irecv() is nonblocking– MPI_Wait() and MPI_Test() can be used to periodically check to see

if the message is ready, and finally wait for it, if desired

Collective Message Passing

• Broadcast– Sends a message from one to all processes in the group

• Scatter– Distributes each element of a data array to a different

process for computation

• Gather– The reverse of scatter…retrieves data elements into an

array from multiple processes

Collective Message Passing w/MPI

MPI_Bcast() Broadcast from root to all other processes

MPI_Gather() Gather values for group of processes

MPI_Scatter() Scatters buffer in parts to group of processes

MPI_Alltoall() Sends data from all processes to all processes

MPI_Reduce() Combine values on all processes to single val

MPI_Reduce_Scatter() Broadcast from root to all other processes

MPI_Bcast() Broadcast from root to all other processes

Message Passing Deadlock

• Deadlock can occur when all critical processes are waiting for messages that never come, or waiting for buffers to clear out so that their own messages can be sent

• Possible Causes– Program/algorithm errors

– Message and buffer sizes

• Solutions– Order operations more carefully

– Use nonblocking operations

– Add debugging output statements to your code to find the problem

Sample PBS Script

sequoia% vi example.pbs #!/bin/bash#PBS -l nodes=4 # Mimosa#PBS –l ncpus=4 # Redwood#PBS –l ncpus=4 # Sequoia#PBS –l cput=0:5:0 # Request 5 minutes of CPU time#PBS –N examplecd $PWDrm *.pbs.[eo]*icc –lmpi –o add_mpi.exe add_mpi.c #Sequoiampiexec -n 4 add_mpi.exe #Sequoia

sequoia % qsub example.pbs37537.sequoia.mcsr.olemiss.edu

PBS: Querying Jobs

MPI Programming Exercises

Hello World

sequential

parallel (w/MPI and PBS)

Add the prime numbers in an Array of numbers

sequential

parallel (w/MPI and PBS)

Log in to sequoia & get workshop filesA. Use secure shell to login from your PC to hpcwoods

ssh [email protected]

B. Use secure shell to from hpcwoods to your training account on sequoia:

ssh tracct1@sequoia

ssh tracct2@sequoia

C. Copy workshop files into your home directory by running: /usr/local/apps/ppro/prepare_mpi_workshop

Examine, compile, and execute hello.c

Examine hello_mpi.c

Examine hello_mpi.c

Add macro to include theheader file for the MPI library calls.

Examine hello_mpi.c

Add function call to initialize

the MPI environment

Examine hello_mpi.c

Add function call find out how many parallel

processes there are.

Examine hello_mpi.c

Add function call to find out which process

this is – the MPI process ID of this process.

Examine hello_mpi.c

Add IF structure so that the manager/boss process can do

one thing, and everyone else (the workers/servants)

can do something else.

Examine hello_mpi.c

All processes, whether manager or worker, must finalize MPI operations.

Compile hello_mpi.c

Why won’t this compile?

You must link to the MPI library.

Compile it.

Run hello_mpi.exe

On 1 CPU

On 2 CPUs

On 4 CPUs

hello_mpi.pbs

hello_mpi.pbs

hello_mpi.pbs

hello_mpi.pbs

hello_mpi.pbs

hello_mpi.pbs

Submit hello_mpi.pbs

Submit hello_mpi.pbs

Submit hello_mpi.pbs

Examine, compile, and execute add_mpi.c

Examine, compile, and execute add_mpi.c

Examine, compile, and execute add_mpi.c

Examine, compile, and execute add_mpi.c

Examine, compile, and execute add_mpi.c

Examine, compile, and execute add_mpi.c

Examine, compile, and execute add_mpi.c

Examine, compile, and execute add_mpi.c

Examine, compile, and execute add_mpi.c

Examine, compile, and execute add_mpi.c

Examine, compile, and execute add_mpi.c

Examine, compile, and execute add_mpi.c

Examine add_mpi.pbs

Examine add_mpi.pbs

Examine add_mpi.pbs

Submit PBS Script: add_mpi.pbs

Examine Output and Errors add_mpi.c

Determine Speedup

Determine Parallel Efficiency

How Could Speedup/Efficiency Improve?

What Happens to ResultsWhen MAXSIZE NotEvenly Divisible by n?

Exercise 1:Change Code to Work When

MAXSIZE is Not EvenlyDivisible by n

Exercise 2:Change Code to Improve Speedup