Introductions to Parallel Programming Using OpenMP

Zhenying Liu, Dr. Barbara ChapmanHigh Performance Computing and Tools group

Computer Science Department

University of Houston

April 7, 2005

Content

• Overview of OpenMP

• Acknowledgement

• OpenMP constructs (5 categories)

• OpenMP exercises

• References

Overview of OpenMP• OpenMP is a set of extensions to Fortran/C/C++

• OpenMP contains compiler directives, library routines and environment variables.

• Available on most single address space machines.

– shared memory systems, including cc-NUMA

• Chip MultiThreading: Chip MultiProcessing (Sun UltraSPARC IV), Simultaneous Multithreading (Intel Xeon)

– not on distributed memory systems, classic MPPs, or PC clusters (yet!)

Shared Memory Architecture

• All processors have access to one global memory

• All processors share the same address space • The system runs a single copy of the OS• Processors communicate by reading/writing to

the global memory• Examples: multiprocessor PCs (Intel P4), Sun

Fire 15K, NEC SX-7, Fujitsu PrimePower, IBM p690, SGI Origin 3000.

Shared Memory Systems (cont)

OpenMP

Pthreads

Processor

M

Processor

M

Processor

M

Interconnect

•••••••••••••

Distributed Memory Systems

MPI

HPF

Clustered of SMPs

MPI

hybrid MPI + OpenMP

OpenMP Usage

• Applications– Applications with intense computational needs– From video games to big science &

engineering

• Programmer Accessibility– From very early programmers in school to

scientists to parallel computing experts

• Available to millions of programmers– In every major (Fortran & C/C++) compiler

OpenMP Syntax

• Most of the constructs in OpenMP are compiler directives or pragmas.– For C and C++, the pragmas take the form:

• #pragma omp construct [clause [clause]…]

– For Fortran, the directives take one of the forms:• C$OMP construct [clause [clause]…]• !$OMP construct [clause [clause]…]• *$OMP construct [clause [clause]…]

• Since the constructs are directives, an OpenMP program can be compiled by compilers that don’t support OpenMP.

OpenMP: Programming Model

• Fork-Join Parallelism:– Master thread spawns a team of threads as needed. – Parallelism is added incrementally: i.e. the sequential

program evolves into a parallel program.

OpenMP:How is OpenMP Typically Used?

• OpenMP is usually used to parallelize loops:– Find your most time consuming loops.

– Split them up between threads.

void main()

{

double Res[1000];

#pragma omp parallel for

for(int i=0;i<1000;i++) {

do_huge_comp(Res[i]);

}

}

void main()

{

double Res[1000];

for(int i=0;i<1000;i++) {

do_huge_comp(Res[i]);

}

}

Split-up this loop between multiple threads

Sequential program Parallel program

OpenMP:How do Threads Interact?

• OpenMP is a shared memory model.– Threads communicate by sharing variables.

• Unintended sharing of data can lead to race conditions:– race condition: when the program’s outcome changes

as the threads are scheduled differently.

• To control race conditions:– Use synchronization to protect data conflicts.

• Synchronization is expensive so:– Change how data is stored to minimize the need for

synchronization.

OpenMP vs. POSIX Threads

• POSIX threads is the other widely used shared programming

API.

• Fairly widely available, usually quite simple to implement on top

of OS kernel threads.

• Lower level of abstraction than OpenMP– library routines only, no directives– more flexible, but harder to implement and maintain– OpenMP can be implemented on top of POSIX threads

• Not much difference in availability– not that many OpenMP C++ implementations– no standard Fortran interface for POSIX threads

Content

• Overview of OpenMP

• Acknowledgement

• OpenMP constructs (5 categories)

• OpenMP exercises

• References

Acknowledgement

• Slides provided by– Tim Mattson and Rudolf Eigenmann, SC ’99– Mark Bull from EPCC

• OpenMP program examples– Lawrence Livermore National Lab– NAS FT parallelization from PGI tutorial

• Dr. Garbey provided us serial codes of Naiver-Stokes

Content

• Overview of OpenMP

• Acknowledgement

• OpenMP constructs (5 categories)

• OpenMP exercises

• References

OpenMP Constructs

• OpenMP’s constructs fall into 5 categories:– Parallel Regions– Worksharing– Data Environment– Synchronization– Runtime functions/environment variables

• OpenMP is basically the same between Fortran and C/C++

OpenMP: Parallel Regions

• You create threads in OpenMP with the “omp parallel” pragma.• For example, To create a 4-thread Parallel region:

• Each thread calls pooh(ID,A) for ID = 0 to 3

double A[1000];

omp_set_num_threads(4);

#pragma omp parallel

{

int ID =omp_get_thread_num();

pooh(ID,A);

}

Each thread redundantly executes the code within the structured block

OpenMP: Work-Sharing Constructs

• The “for” Work-Sharing construct splits up loop iterations among the threads in a team

#pragma omp parallel

#pragma omp for

for (I=0;I<N;I++){

NEAT_STUFF(I);

}

By default, there is a barrier at the end of

the “omp for”. Use the “nowait” clause to

turn off the barrier.

Work Sharing ConstructsA motivating examplefor(i=0;I<N;i++) { a[i] = a[i] + b[i];}

#pragma omp parallel

{

int id, i, Nthrds, istart, iend;

id = omp_get_thread_num();

Nthrds = omp_get_num_threads();

istart = id * N / Nthrds;

iend = (id+1) * N / Nthrds;

for(i=istart;I<iend;i++) {a[i]=a[i]+b[i];}

}

#pragma omp parallel

#pragma omp for schedule(static)

for(i=0;I<N;i++) { a[i]=a[i]+b[i];}

OpenMP parallel region and a work-sharing for construct

Sequential code

OpenMP Parallel Region

OpenMP Parallel Region and a work-sharing for construct

OpenMP For Construct:The Schedule Clause

• The schedule clause effects how loop iterations are mapped onto threads

schedule(static [,chunk])• Deal-out blocks of iterations of size “chunk” to each thread.

schedule(dynamic[,chunk])• Each thread grabs “chunk” iterations off a queue until all

iterations have been handled.

schedule(guided[,chunk])• Threads dynamically grab blocks of iterations. The size of the

block starts large and shrinks down to size “chunk” as the calculation proceeds.

schedule(runtime)• Schedule and chunk size taken from the OMP_SCHEDULE

environment variable.

OpenMP: Work-Sharing Constructs

• The Sections work-sharing construct gives a different structured block to each thread.

#pragma omp parallel

#pragma omp sections

{

X_calculation();

#pragma omp section

y_calculation();

#pragma omp section

z_calculation();

}

By default, there is a barrier at the end of the “omp sections”. Use the “nowait” clause to turn off the barrier.

Data Environment:Changing Storage Attributes

• One can selectively change storage attributes constructs using the following clauses*– SHARED– PRIVATE– FIRSTPRIVATE– THREADPRIVATE

• The value of a private inside a parallel loop can be transmitted to a global value outside the loop with:– LASTPRIVATE

• The default status can be modified with:– DEFAULT (PRIVATE | SHARED | NONE)

* All data clauses apply to parallel regions and worksharing constructs except “shared” which only applies to parallel regions.

Data Environment:Default Storage Attributes

• Shared Memory programming model:– Most variables are shared by default

• Global variables are SHARED among threads– Fortran: COMMON blocks, SAVE variables, MODULE

variables– C: File scope variables, static

• But not everything is shared...– Stack variables in sub-programs called from parallel

regions are PRIVATE– Automatic variables within a statement block are

PRIVATE.

Private Clause• private(var) creates a local copy of var for

each thread.

– The value is uninitialized

– Private copy is not storage associated with the original

void wrong(){int IS = 0;

#pragma parallel for private(IS)for(int J=1;J<1000;J++)

IS = IS + J;printf(“%i”, IS);

}

OpenMP: Reduction

• Another clause that effects the way variables are shared:– reduction (op : list)

• The variables in “list” must be shared in the enclosing parallel region.

• Inside a parallel or a worksharing construct:– A local copy of each list variable is made and

initialized depending on the “op” (e.g. 0 for “+”)– pair wise “op” is updated on the local value– Local copies are reduced into a single global copy at

the end of the construct.

OpenMP: An Reduction Example

#include <omp.h>#define NUM_THREADS 2void main (){

int i;double ZZ, func(), sum=0.0;omp_set_num_threads(NUM_THREADS)#pragma omp parallel for reduction(+:sum) private(ZZ)for (i=0; i< 1000; i++){

ZZ = func(i);sum = sum + ZZ;

}}

OpenMP: Synchronization

• OpenMP has the following constructs to support synchronization:– barrier– critical section– atomic– flush– ordered– single– master

Critical and Atomic

• Only one thread at a time can enter a critical section

C$OMP PARALLEL DO PRIVATE(B)C$OMP& SHARED(RES) DO 100 I=1,NITERS

B = DOIT(I)C$OMP CRITICAL

CALL CONSUME (B, RES)C$OMP END CRITICAL100 CONTINUE

C$OMP PARALLEL PRIVATE(B)B = DOIT(I)

C$OMP ATOMICX = X + B

C$OMP END PARALLEL

• Atomic is a special case of a critical section that can be used for certain simple statements:

Master directive• The master construct denotes a structured block that

is only executed by the master thread. The other threads just skip it (no implied barriers or flushes).

#pragma omp parallel private (tmp)

{

do_many_things();

#pragma omp master

{ exchange_boundaries(); }

#pragma barrier

do_many_other_things();

}

Single directive

• The single construct denotes a block of code that is executed by only one thread.

• A barrier and a flush are implied at the end of the single block.

#pragma omp parallel private (tmp){

do_many_things();#pragma omp single{ exchange_boundaries(); }do_many_other_things();

}

OpenMP: Library routines• Lock routines

– omp_init_lock(), omp_set_lock(), omp_unset_lock(), omp_test_lock()

• Runtime environment routines:– Modify/Check the number of threads

• omp_set_num_threads(), omp_get_num_threads(), omp_get_thread_num(), omp_get_max_threads()

– Turn on/off nesting and dynamic mode• omp_set_nested(), omp_set_dynamic(), omp_get_nested(),

omp_get_dynamic()

– Are we in a parallel region?• omp_in_parallel()

– How many processors in the system?• omp_num_procs()

OpenMP: Environment Variables

• OMP_NUM_THREADS– bsh:

• export OMP_NUM_THREADS=2

– csh:• setenv OMP_NUM_THREADS 4

Content

• Overview of OpenMP

• Acknowledgement

• OpenMP constructs (5 categories)

• OpenMP exercises

• References

#include <omp.h>main () {int nthreads, tid;/* Fork a team of threads giving them their own copies of variables */#pragma omp parallel private(nthreads, tid) { /* Obtain thread number */ tid = omp_get_thread_num(); printf("Hello World from thread = %d\n", tid); /* Only master thread does this */ if (tid == 0) { nthreads = omp_get_num_threads(); printf("Number of threads = %d\n", nthreads); } } /* All threads join master thread and disband */}

1. Hello World!

Example Code - Pthread Creation and Termination #include <pthread.h> #include <stdio.h> #define NUM_THREADS 5 void *PrintHello(void *threadid) { printf("\n%d: Hello World!\n", threadid); pthread_exit(NULL); }

int main (int argc, char *argv[]) { pthread_t threads[NUM_THREADS]; int rc, t; for(t=0; t<NUM_THREADS; t++) { printf("Creating thread %d\n", t); rc = pthread_create(&threads[t], NULL, PrintHello, (void *)t); if (rc) { printf("ERROR; return code from pthread_create() is %d\n", rc); exit(-1); } } pthread_exit(NULL); }

PROGRAM REDUCTION

INTEGER I, N REAL A(100), B(100), SUM

! Some initializations N = 100 DO I = 1, N A(I) = I *1.0 B(I) = A(I) ENDDO SUM = 0.0

!$OMP PARALLEL DO REDUCTION(+:SUM) DO I = 1, N SUM = SUM + (A(I) * B(I)) ENDDO

PRINT *, ' Sum = ', SUM END

2. Parallel Loop Reduction

3. Matrix-vector multiply using a parallel loop and critical directive

/*** Spawn a parallel region explicitly scoping all variables ***/#pragma omp parallel shared(a,b,c,nthreads,chunk) private(tid,i,j,k){ #pragma omp for schedule (static, chunk) for (i=0; i<NRA; i++) { printf("thread=%d did row=%d\n",tid,i); for(j=0; j<NCB; j++) for (k=0; k<NCA; k++) c[i][j] += a[i][k] * b[k][j]; }}

Steps of Parallelization using OpenMP: An Example from a PGI Tutorial

1. Compile a code with the option to enable a profiler

2. Run the code and check if the results are correct

3. Find out the most time-consuming part of the code via the profiler information

4. Parallelize the time-consuming part5. Repeat above steps until you get

reasonable speedup

How to Use a Profiler

• PGI compiler– pgf90 -fast -Minfo -Mprof=func fftpde.F -o

fftpde (function level)• -Mprof=lines (line level)• -mp for compiling OpenMP codes

– pgprof pgprof.out (show the profiler result)

• Pathscale compiler– pathf90 -Ofast -pg Fftpde.F -o Fftpde

• pathprof Fftpde|more

do k=1,n3 do j=1,n2 do i=1,n1 z(i)=cmplx(x1real(i,j,k),x1imag(i,j,k)) end do call fft(z,inverse,w,n1,m1) do i=1,n1 x1real(i,j,k)=real(z(i)) x1imag(i,j,k)=aimag(z(i)) end do end doend do

The most time-consuming loop in Fftpde.F:

!$OMP PARALLEL PRIVATE(Z)!$OMP DOdo k=1,n3 do j=1,n2 do i=1,n1 z(i)=cmplx(x1real(i,j,k),x1imag(i,j,k)) end do call fft(z,inverse,w,n1,m1) do i=1,n1 x1real(i,j,k)=real(z(i)) x1imag(i,j,k)=aimag(z(i)) end do end doend do!$OMP END PARALLEL

The OpenMP version of this loop in Fftpde_1.F:

NEXT: compare the 1 and 2 processor profiles after adding OpenMP to this loop

Parallelizing the Reminder of Fftpde.F

1. The DO 130 loop near line 64 (fftpde_2.F)

2. The DO 190 loop near line 115 (fftpde_3.F)

3. 3) The DO 220 loop near line 139 (fftpde_4.F)

4. 4) The DO 250 loop near line 155 (fftpde_5.F)

!$OMP PARALLEL PRIVATE(KK,KL,T1,T2,IK)!$OMP DO DO 130 K = 1, N3 KK = K - 1 KL = KK T1 = S T2 = ANCC Find starting seed T1 for this KK using the binary rule for exponentiation.C DO 110 I = 1, 100 IK = KK / 2 IF (2 * IK .NE. KK) T2 = RANDLC (T1, T2) IF (IK .EQ. 0) GOTO 120 T2 = RANDLC (T2, T2) KK = IK 110 CONTINUECC Compute 2 * NQ pseudorandom numbers.C 120 continue CALL VRANLC (N1*N2, T2, aa, x1real(1,1,k)) CALL VRANLC (N1*N2, T2, aa, x1imag(1,1,k)) 130 CONTINUE!$OMP END PARALLEL

1. Parallelize the DO 130 loop in Fftpde_2.F

!$OMP PARALLEL PRIVATE(K1,J1,JK,I1)!$OMP DO DO 190 K = 1, N3 K1 = K - 1 IF (K .GT. N32) K1 = K1 - N3C DO 180 J = 1, N2 J1 = J - 1 IF (J .GT. N22) J1 = J1 - N2 JK = J1 ** 2 + K1 ** 2C DO 170 I = 1, N1 I1 = I - 1 IF (I .GT. N12) I1 = I1 - N1 X3(I,J,K) = EXP (AP * (I1 ** 2 + JK)) 170 CONTINUEC 180 CONTINUE 190 CONTINUE!$OMP END PARALLEL

2. Parallelize the DO 190 loop in Fftpde_3.F

3. Parallelize the DO 220 loop in Fftpde_4.F

!$OMP PARALLEL PRIVATE(T1)!$OMP DO DO 220 K = 1, N3 DO 210 J = 1, N2 DO 200 I = 1, N1 T1 = X3(I,J,K) ** KT X2real(I,J,K) = T1 * X1real(I,J,K) X2imag(I,J,K) = T1 * X1imag(I,J,K) 200 CONTINUE 210 CONTINUE 220 CONTINUE!$OMP END PARALLEL

4. Parallelize the DO 250 loop in Fftpde_5.F

!$OMP PARALLEL!$OMP DO DO 250 K = 1, N3 DO 240 J = 1, N2 DO 230 I = 1, N1 X2real(I,J,K) = RN * X2real(I,J,K) X2imag(I,J,K) = RN * X2imag(I,J,K) 230 CONTINUE 240 CONTINUE 250 CONTINUE!$OMP END PARALLEL

Conclusion

• OpenMP is successful in small-to-medium SMP systems

• Multiple cores/CPUs dominate the future computer architectures; OpenMP would be the major parallel programming language in these architectures.

• Simple: everybody can learn it in 2 weeks• Not so simple: Don’t stop learning! keep learning

it for better performance

Some Buggy Codes

#pragma omp parallel for shared(a,b,c,chunk) private(i,tid) schedule(static,chunk) { tid = omp_get_thread_num(); for (i=0; i < N; i++) { c[i] = a[i] + b[i]; printf("tid= %d i= %d c[i]= %f\n", tid, i, c[i]); } } /* end of parallel for construct */

}

Content

• Overview of OpenMP

• Acknowledgement

• OpenMP constructs (5 categories)

• OpenMP exercises

• References

References

• OpenMP Official Website:– www.openmp.org

• OpenMP 2.5 Specifications• An OpenMP book

– Rohit Chandra, “Parallel Programming in OpenMP”. Morgan Kaufmann Publishers.

• Compunity– The community of OpenMP researchers and developers in

academia and industry – http://www.compunity.org/

• Conference papers:– WOMPAT, EWOMP, WOMPEI, IWOMP

• http://www.nic.uoregon.edu/iwomp2005/index.html#program

Exercises

• cp /tmp/omp_examples.tar.gz ~/• tar –xzvf omp_examples.tar.gz• (marvin:) use pathscale; (medusa:) use pgi

– pathf90; pathcc -mp -Ofast– pgf90; pgcc -mp -fast

• Compile(make) and run the codes in– LLNL_C, LLNL_F, and FFT

• There is a README in each subdirectory

• Set the number of threads before running– Echo $SHELL– export OMP_NUM_THREADS=2 (for bsh)– setenv OMP_NUM_THREADS=2 (for csh)

OpenMP Compilers and Platforms

• Fujitsu/Lahey Fortran, C and C++ – Intel Linux Systems – Fujitsu Solaris Systems

• HP HP-UX PA-RISC/Itanium , HP Tru64 Unix – Fortran/C/C++

• IBM XL Fortran and C from IBM – IBM AIX Systems

• Intel C++ and Fortran Compilers from Intel – Intel IA32 Linux/Windows Systems – Intel Itanium-based Linux/Windows Systems

• Guide Fortran and C/C++ from Intel's KAI Software Lab – Intel Linux/Windows Systems

• PGF77 and PGF90 Compilers from The Portland Group, Inc. (PGI) – Intel Linux/Solaris/Windows/NT Systems

Compilers and Platforms

• SGI MIPSpro 7.4 Compilers – SGI IRIX Systems

• Sun Microsystems Sun ONE Studio, Compiler Collection, Fortran 95, C, and C++

– Sun Solaris Platforms • VAST from Veridian Pacific-Sierra Research

– IBM AIX Systems – Intel IA32 Linux/Windows/NT Systems – SGI IRIX Systems

– Sun Solaris Systems • PATHSCALE EKOPATH™ COMPILER SUITE FOR AMD64 and

EM64T, Fortran, C, C++– 64-bit Linux

• Microsoft Visual Studio 2005 (Visual C++)– Windows

Parallelize: Win32 API, PIvoid main ()

{

double pi; int i;

DWORD threadID;

int threadArg[NUM_THREADS];

for(i=0; i<NUM_THREADS; i++) threadArg[i] = i+1;

InitializeCriticalSection(&hUpdateMutex);

for (i=0; i<NUM_THREADS; i++){

thread_handles[i] = CreateThread(0, 0, (LPTHREAD_START_ROUTINE) Pi, &threadArg[i], 0, &threadID);

}

WaitForMultipleObjects(NUM_THREADS,

thread_handles, TRUE,INFINITE);

pi = global_sum * step;

printf(" pi is %f \n",pi);

}

#include <windows.h>

#define NUM_THREADS 2

HANDLE thread_handles[NUM_THREADS];

CRITICAL_SECTION hUpdateMutex;

static long num_steps = 100000;

double step;

double global_sum = 0.0;

void Pi (void *arg)

{

int i, start;

double x, sum = 0.0;

start = *(int *) arg;

step = 1.0/(double) num_steps;

for (i=start;i<= num_steps; i=i+NUM_THREADS){

x = (i-0.5)*step;

sum = sum + 4.0/(1.0+x*x);

}

EnterCriticalSection(&hUpdateMutex);

global_sum += sum;

LeaveCriticalSection(&hUpdateMutex);

} Doubles code size!

Solution: Keep it simple

• Threads libraries:– Pro: Programmer has control over everything– Con: Programmer must control everything

Programmers

scared away

Full

control

Increased

complexity

Sometimes a simple evolutionary approach is better

PI Program: an examplestatic long num_steps = 100000;

double step;

void main ()

{ int i; double x, pi, sum = 0.0;

step = 1.0/(double) num_steps;

for (i=1;i<= num_steps; i++){

x = (i-0.5)*step;

sum = sum + 4.0/(1.0+x*x);

}

pi = step * sum;

}

OpenMP PI Program: Parallel Region example (SPMD Program)

#include <omp.h>static long num_steps = 100000; double step;#define NUM_THREADS 2void main (){ int i; double x, pi, sum[NUM_THREADS]; step = 1.0/(double) num_steps; omp_set_num_threads(NUM_THREADS); #pragma omp parallel { double x; int id; id = omp_get_thread_num(); for (i=id, sum[id]=0.0;i< num_steps; i=i+NUM_THREADS){ x = (i+0.5)*step;

sum[id] += 4.0/(1.0+x*x);}

} for(i=0, pi=0.0;i<NUM_THREADS;i++) pi += sum[i] * step;}

SPMD Programs:

Each thread runs the same code with the thread ID

selecting any thread specific behavior.

OpenMP PI Program:Work Sharing Construct

#include <omp.h>

static long num_steps = 100000; double step;

#define NUM_THREADS 2

void main ()

{ int i; double x, pi, sum[NUM_THREADS]; step = 1.0/(double) num_steps; omp_set_num_threads(NUM_THREADS) #pragma omp parallel { double x; int id; id = omp_get_thread_num(); sum[id] = 0; #pragma omp for for (i=id;i< num_steps; i++){ x = (i+0.5)*step;

sum[id] += 4.0/(1.0+x*x); } } for(i=0, pi=0.0;i<NUM_THREADS;i++)pi += sum[i] * step;}

OpenMP PI Program:Private Clause and a Critical Section

#include <omp.h>

static long num_steps = 100000; double step;

#define NUM_THREADS 2

void main ()

{ int i, id; double x, sum, pi=0.0;step = 1.0/(double) num_steps;omp_set_num_threads(NUM_THREADS);#pragma omp parallel private (x, sum){ id = omp_get_thread_num(); for (i=id,sum=0.0;i< num_steps;i=i+NUM_THREADS){

x = (i+0.5)*step;sum += 4.0/(1.0+x*x);

} #pragma omp critical pi += sum}

}

Note: We didn’t need to create an array to hold local sums or clutter the code with explicit declarations of “x” and “sum”.

OpenMP PI Program:Parallel For with a Reduction

#include <omp.h>

static long num_steps = 100000; double step;

#define NUM_THREADS 2

void main ()

{ int i; double x, pi, sum = 0.0;step = 1.0/(double) num_steps;

omp_set_num_threads(NUM_THREADS);

#pragma omp parallel for reduction(+:sum) private(x)

for (i=1;i<= num_steps; i++){x = (i-0.5)*step;sum = sum + 4.0/(1.0+x*x);

}

pi = step * sum;

}

OpenMP adds 2 to 4

lines of code