Introduction to Parallel Computing

1
Introduction to Parallel Computing
Presentation
Martin CumaCenter for High Performance Computing
University of Utah mcuma_at_chpc.utah.edu
April 12, 2007
http//www.chpc.utah.edu
2
Overview

• Types of parallel computers.
• Parallel programming options.
• How to write parallel applications.
• How to compile.
• How to debug/profile.
• Summary, future expansion.

3
Parallel architectures

• Single processor
• SISD single instruction single data.
• Multiple processors
• SIMD - single instruction multiple data.
• MIMD multiple instruction multiple data.
• Shared Memory
• Distributed Memory

4
Shared memory

• All processors have access to local memory
• Simpler programming
• Concurrent memory access
• More specialized hardware
• CHPC Arches dual and quad CPU nodes

5
Distributed memory

• Process has access only to its local memory
• Data between processes must be communicated
• More complex programming
• Cheap commodity hardware
• CHPC Linux clusters (Arches)

6
Parallel programming options

• Shared Memory
• Threads POSIX Pthreads, OpenMP
• Message passing
• Distributed Memory
• Message passing libraries
• Vendor specific non portable
• General MPI, PVM

7
OpenMP basics

• Compiler directives to parallelize
• Fortran source code comments
• !omp parallel/!omp end parallel
• C/C - pragmas
• pragma omp parallel
• Small set of subroutines
• Degree of parallelism specification
• OMP_NUM_THREADS or omp_set_num_threads(INTEGER n)

8
MPI Basics

• Communication library
• Language bindings
• C/C - int MPI_Init(int argv, char argc)
• Fortran - MPI_Init(INTEGER ierr)
• Quite complex (100 subroutines)
• but only small number used frequently
• Fixed number of parallel nodes

9
MPI vs. OpenMP

• Easy to code
• Fast data exchange
• Memory access (thread safety)
• Limited usability
• Limited users influence on parallel execution

• Complex to code
• Slow data communication
• Ported to many architectures
• Many tune-up options for parallel execution

10
Program example

• saxpy vector addition
• simple loop, no cross-dependence, easy to
  parallelize
• subroutine saxpy_serial(z, a, x, y, n)
• integer i, n
• real z(n), a, x(n), y(n)
• do i1, n
• z(i) ax(i) y(i)
• enddo
• return

11
OpenMP program example

• subroutine saxpy_parallel_omp(z, a, x, y, n)
• integer i, n
• real z(n), a, x(n), y(n)
• !omp parallel do
• do i1, n
• z(i) ax(i) y(i)
• enddo
• return
• setenv OMP_NUM_THREADS 4

12
MPI program example

• subroutine saxpy_parallel_mpi(z, a, x, y, n)
• integer i, n, ierr, my_rank, nodes, i_st, i_end
• real z(n), a, x(n), y(n)
• call MPI_Init(ierr)
• call MPI_Comm_rank(MPI_COMM_WORLD,my_rank,ierr)
• call MPI_Comm_size(MPI_COMM_WORLD,nodes,ierr)
• i_st n/nodesmy_rank1
• i_end n/nodes(my_rank1)
• do ii_st, i_end
• z(i) ax(i) y(i)
• enddo
• call MPI_Finalize(ierr)
• return

z(i) operation on 4 nodes
z(1 n/4)
z(n/41 2n/4)
z(2n/41 3n/4)
z(3n/41 n)
13
MPI program example

• Result on the first node
• include "mpif.h"
• integer status(MPI_STATUS_SIZE)
• if (my_rank .eq. 0 ) then
• do j 1, nodes-1
• do i n/nodesj1, n/nodes(j1)
• call MPI_Recv(z(i),1,MPI_REAL,j,0,MPI_COMM_W
  ORLD,
• status,ierr)
• enddo
• enddo
• else
• do ii_st, i_end
• call MPI_Send(z(i),1,MPI_REAL,0,0,MPI_COMM_WORL
  D,ierr)
• enddo
• endif

14
MPI program example
zi(i)
Node 1

• Collective communication
• real zi(n)
• j 1
• do ii_st, i_endzi(j) ax(i) y(i)j j 1
• enddo
• call MPI_Gather(zi,n/nodes,MPI_REAL,z,n/nodes,MPI_
  REAL,
• 0,MPI_COMM_WORLD,ierr)
• Result on all nodes
• call MPI_AllGather(zi,n/nodes,MPI_REAL,z,n/nodes,
• MPI_REAL,MPI_COMM_WORLD,ierr)

zi(i)
Node 2
zi(i)
Node 3
zi(i)
Node 4
z(i)
Receive data
Send data
No root process
15
Arches - login

• First log into Arches
• ssh delicatearch.chpc.utah.edu - Myrinet
• ssh marchingmen.chpc.utah.edu - Ethernet
• ssh tunnelarch.chpc.utah.edu - Ethernet
• ssh landscapearch.chpc.utah.edu Ethernet,
  Myrinet
• ssh sanddunearch.chpc.utah.edu - InfiniBand
• Then submit a job to get compute nodes
• qsub I l nodes2ppn4,walltime10000
• qsub script.pbs
• Useful scheduler commands
• qsub submit a job
• qdel delete a job
• showq show job queue

16
Security Policies

• No clear text passwords use ssh and scp
• You may not share your account under any
  circumstances
• Dont leave your terminal unattended while logged
  into your account
• Do not introduce classified or sensitive work
  onto CHPC systems
• Use a good password and protect it

17
Security Policies

• Do not try to break passwords, tamper with files
  etc.
• Do not distribute or copy privileged data or
  software
• Report suspicians to CHPC (security_at_chpc.utah.edu)
  
• Please see http//www.chpc.utah.edu/docs/policies/
  security.html for more details

18
Compilation - OpenMP

• Arches
• Supported by Portland Group and Pathscale
  compilers, mp switch
• e.g. pgf77 mp source.f o program.exe
• Dual-processor and quad-processor (Sanddunearch)
  nodes
• Further references
• CHPC website
• Portland Group website
• http//www.pgroup.com/doc/index.htm
• Pathscale website
• http//www.pathscale.com/docs.html

19
Compilation - MPI

• Different implementations on different machines
• MPI code is portable on all CHPC systems

20
Compilation - MPI

• Arches MPICH, MPICH2, MVAPICH, MVAPICH2
• /MPI-path/bin/mpixx source.x o program.exe
• MPI-path location of the distribution
• /uufs/arches/sys/pkg/mpich/std TCP-IP
• /uufs/UUFSCELL/sys/pkg/mpich-mx/std Myrinet
• /uufs/arches/sys/pkg/mpich2/std MPICH2 TCP-IP
• /uufs/sanddunearch.arches/sys/pkg/mvapich/std
  MVAPICH InfiniBand
• /uufs/sanddunearch.arches/sys/pkg/mvapich2/std
  MVAPICH2 InfiniBand
• must specify full path to mpixx
  (/MPI-path/bin) or add this path to PATH
  environment variable

21
Running a parallel job Arches

• MPICH Interactive batch
• qsub I l nodes2ppn2,walltime10000
• wait for prompt
• /MPI-path/bin/mpirun np 4 machinefile
  PBS_NODEFILE program.exe
• /MPI-path/bin/mpirun_rsh rsh np 4
  machinefile PBS_NODEFILE program.exe
• MPICH Batch
• PBS l nodes2ppn2,walltime10000
• OpenMP Batch
• PBS l nodes1ppn2,walltime10000
• setenv OMP_NUM_THREADS 2
• program.exe

22
Running a parallel job Arches

• MPICH2, MVAPICH2 Interactive batch
• qsub I l nodes2ppn2,walltime10000
• wait for prompt
• mpdboot n 2 r /usr/bin/rsh f PBS_NODEFILE
• mpdtrace
• da001
• da002
• mpiexec n 4 ./program.exe
• NOTE on sanddunearch, use ppn4

23
Debuggers vendor supplied

• Arches, desktops
• gdb pgdbg pathdb, idb
• How to use
• http//www.chpc.utah.edu/docs/manuals/software/par
  _devel.html

24
Debuggers - parallel

• Totalview
• Arches, desktops
• Graphical, multi-process debugger
• How to run it
• compile with g flag
• Details
• http//www.chpc.utah.edu/docs/manuals/software/to
  talview.html
• Further information
• http//www.totalviewtech.com/alltvdocs.htm

25
Debuggers parallel cont.
Process view
Source code view
Data inspection
26
Profilers - parallel

• Arches
• serial gprof, pgprof
• parallel TAU/Vampir

27
Profilers - parallel
28
Parallel libraries

• Arches
• ScaLAPACK, PETSc, NAG, FFTW
• http//www.chpc.utah.edu/docs/manuals/software/mat
  _l.html

29
Summary

• Shared vs. Distributed memory
• OpenMP
• limited on Arches
• Simple parallelization
• MPI
• Arches
• Must use communication
• http//www.chpc.utah.edu/docs/presentations/intro_
  par

30
References

• OpenMP
• http//www.openmp.org/
• Chandra, Dagum, Kohr, - Parallel Programming in
  OpenMP
• MPI
• http//www-unix.mcs.anl.gov/mpi/
• Pacheco - Parallel Programming with MPI
• Gropp, Lusk, Skjellum - Using MPI 1, 2

31
Future Presentations

• Introduction to MPI
• Introduction to OpenMP
• Debugging with Totalview
• Profiling with TAU/Vampir
• Intermediate MPI and MPI-IO
• Mathematical Libraries at the CHPC