Nvidia CUDA Toolkit

From UFRC
Revision as of 13:02, 23 September 2013 by Taylor (talk | contribs)
Jump to navigation Jump to search

Description

cuda website  
CUDA™ is a parallel computing platform and programming model invented by NVIDIA. It enables dramatic increases in computing performance by harnessing the power of the graphics processing unit (GPU). With millions of CUDA-enabled GPUs sold to date, software developers, scientists and researchers are finding broad-ranging uses for GPU computing with CUDA.

Required Modules

cuda

System Variables

  • HPC_{{#uppercase:cuda}}_DIR
  • HPC_{{#uppercase:cuda}}_BIN
  • HPC_{{#uppercase:cuda}}_INC
  • HPC_{{#uppercase:cuda}}_LIB

Nvidia GPUs

Research Computing has a significant investment in GPU-enabled servers. Each supports from two to eight Nvidia GPUs that range from the earlier S1070s to the more recent Kepler K20s (see table below).

GPU Quantity Host Architecture Host Memory Host Interconnect Notes
Servers Processor Cores Memory GPUs IB Notes
4 1 Intel E5462 (2.8 GHz) 4 16 GB 2 DDR 1
4 1 Intel E5675 (3.0GHz) 6 24 GB 2 QDR 3,4
10 2 Intel E5-2643 (3.3 GHz) 8 64 GB 4 FDR 2
16 2 AMD Opteron 6220 (3.0 GHz) 16 32 GB 2 QDR 2
1 2 Intel E5620 (2.4GHz) 8 24 GB 8 N/A 3
  1. S1070 GPUs, 4GB Device RAM
  2. M2090 GPUs, 6GB Device RAM
  3. M2070 GPUs, 6GB Device RAM
  4. K20c GPUs, 5GB Device RAM

PBS Script Examples

See the Nvidia CUDA Toolkit_PBS page for cuda PBS script examples.

Usage Policy

Interactive Use

If you need interactive access to a gpu for development and testing you may do so by requesting an interactive session through the batch system.

In order to gain interactive access to a GPU server you should run similar to the one that follows. 

qsub -I -l nodes=1:gpus=1:tesla,walltime=01:00:00 -q gpu

To gain access to one of the Fermi-class GPUs, you can make a similar request but specify the "fermi" attribute in your resource request as below. 

qsub -I -l nodes=1:gpus=1:fermi,walltime=01:00:00 -q gpu

If a gpu is available, you will get a prompt on one of the nodes within a minute or two. Otherwise, you will have to wait or try another time. If you choose to wait, you will be connected when a gpu is available. The default walltime limit for the gpu queue is 10 minutes. You should request the amount of time you need but be sure to log out and end your session when you are finished so that the GPU will be available to others.

If your work needs both GPUs attached to the same node, you would run the following command instead. 

qsub -I -l nodes=1:gpus=2,walltime=01:00:00 -q gpu

If you need to request a particular machine, say tesla1, you would use the following qsub command. 

qsub -I -l nodes=tesla1:gpus=1,walltime=01:00:00 -q gpu

Batch Jobs

The process is much the same for batch jobs. To access a node with an M2090, you can add the following to your submission script. 

#PBS -q gpu
#PBS -l nodes=1:gpus=1:M2090
#PBS -l walltime=1:00:00

To access a node with an M2070 GPU, you can add the following to your submission script. 

#PBS -q gpu
#PBS -l nodes=1:gpus=1:m2070
#PBS -l walltime=1:00:00

Exclusive Mode

The GPUs are configured to run in exclusive mode. This means that the gpu driver will only allow one process at a time to access the GPU. If GPU 0 is in use and your application tries to use it, it will simply block. If your application does not call cudaSetDevice(), the CUDA runtime should assign it to a free GPU. Since everyone will be accessing the GPUs through the batch system, there should be no over-subscription of the GPUs.

Performance

Environment

For CUDA development please load the "cuda" module. Doing so will ensure that your environment is set up correctly for the use of the CUDA compiler, header files, and libraries. 

$ module spider cuda
Rebuilding cache, please wait ... (not written to file) done

    Description:
      NVIDIA CUDA Toolkit

     Versions:
        cuda/4.2
        cuda/5.5

$ module load cuda/5.5

$ which nvcc
/opt/cuda/5.5/bin/nvcc

$ printenv | grep CUDA
HPC_CUDA_LIB=/opt/cuda/5.5/lib64
HPC_CUDA_DIR=/opt/cuda/5.5
HPC_CUDA_BIN=/opt/cuda/5.5/bin
HPC_CUDA_INC=/opt/cuda/5.5/include

Example: CUBLAS vs. Optimized BLAS

Note that double-precision linear algebra is a less than ideal application for the GPUs. Still, it is a functional example of using one of the available CUDA runtime libraries.

cuda_bm.c

{{#fileAnchor: cuda_bm.c}} Download raw source of the [{{#fileLink: cuda_bm.c}} cuda_bm.c] 

#include <stdio.h>
#include <stdlib.h>
#include <mkl.h>
#include <math.h>
#include <time.h>
#include <cublas.h>

void resuse(char *str);

double timeDiff( struct timespec *t1, struct timespec *t2)
{
  double T1, T2;
  T2 = (double)t2->tv_sec + (double)t2->tv_nsec / 1.0e9;
  T1 = (double)t1->tv_sec - (double)t1->tv_nsec / 1.0e9;
  return(T2 - T1);
}

main()
{
  int       dim = 9100;
  int       i,j,k;
  int       status;

  double    *psa, *psb, *psc;
  double    *sap, *sbp, *scp;
  double    *pda, *pdb, *pdc;
  double    *dap, *dbp, *dcp;

  double alpha   = 1.0;
  double beta    = 0.0;
  double gflops  = 0.0;
  float  deltaT  = 0.0;
  double gflopCnt = 2.0 * dim * dim * dim / 1.0e9;
  struct timespec t1;
  struct timespec t2;

  int       ptime();

  pda = NULL;
  pdb = NULL;
  pdc = NULL;
  psa = (double *) malloc(dim * dim * sizeof(*psa) );
  psb = (double *) malloc(dim * dim * sizeof(*psb) );
  psc = (double *) malloc(dim * dim * sizeof(*psc) );

  printf("Initializing Matrices...");
  clock_gettime(CLOCK_MONOTONIC, &t1); 
  sap = psa;
  sbp = psb;
  scp = psc;
  for (i = 0; i < dim; i++)
    for (j = 0; j < dim; j++) {
      *sap++ = 1.0;
      *sbp++ = 1.0;
      *scp++ = 0.0;
    }
  clock_gettime(CLOCK_MONOTONIC, &t2); 
  deltaT = timeDiff(&t1, &t2);
  printf("Done.    Elapsed Time = %6.4f secs\n", deltaT);
  fflush(stdout);

  printf("Starting parallel DGEMM...");
  fflush(stdout);
  clock_gettime(CLOCK_MONOTONIC, &t1); 
  cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, dim, dim, dim, alpha, psa, dim, psb, dim, beta, psc, dim);
  clock_gettime(CLOCK_MONOTONIC, &t2); 
  deltaT = timeDiff(&t1, &t2);
  printf("Done.  Elapsed Time = %6.4f secs\n", deltaT);
  printf("                                 ");
  printf("GFlOP Rate   = %8.4f\n", gflopCnt/deltaT);

  if ( (float) dim - psc[0] > 1.0e-5  ||
       (float) dim - psc[dim*dim-1] > 1.0e-5 ) {
      printf("Error: Incorrect Results!\n");
      printf("C[%2d,%2d] = %10.4f\n", 1,1,psc[0]); 
      printf("C[%2d,%2d] = %10.4f\n", dim,dim,psc[dim*dim-1]); 
  }

  /* Initialize CUDA */
  printf("Initializing CUDA...");
  status = cublasInit();
  if (status != CUBLAS_STATUS_SUCCESS) {
      fprintf (stderr, "!!!! CUBLAS initialization error\n");
      return EXIT_FAILURE;
  }
  printf("Done.\n");

  /* Re-initialize the matrices */
  printf("Re-initializing Matrices...");
  clock_gettime(CLOCK_MONOTONIC, &t1); 
  sap = psa;
  sbp = psb;
  scp = psc;
  for (i = 0; i < dim; i++) {
    for (j = 0; j < dim; j++) {
      *sap++ = 1.0;
      *sbp++ = 1.0;
      *scp++ = 0.0;
    }
  }
  clock_gettime(CLOCK_MONOTONIC, &t2); 
  deltaT = timeDiff(&t1, &t2);
  printf("Done.  Elapsed Time = %6.4f secs\n", deltaT);
  fflush(stdout);

  /* Allocate device memory for the matrices */
  printf("Starting CUDA DGEMM...");
  clock_gettime(CLOCK_MONOTONIC, &t1); 

  status = cublasAlloc(dim*dim, sizeof(*pda), (void**) &pda);
  if (status != CUBLAS_STATUS_SUCCESS) {
      fprintf (stderr, "!!!! device memory allocation error (A)\n");
      return EXIT_FAILURE;
  }

  status = cublasAlloc(dim*dim, sizeof(*pdb), (void**) &pdb);
  if (status != CUBLAS_STATUS_SUCCESS) {
      fprintf (stderr, "!!!! device memory allocation error (B)\n");
      return EXIT_FAILURE;
  }

  status = cublasAlloc(dim*dim, sizeof(*pdc), (void**) &pdc);
  if (status != CUBLAS_STATUS_SUCCESS) {
      fprintf (stderr, "!!!! device memory allocation error (C)\n");
      return EXIT_FAILURE;
  }

  /* Initialize the device matrices with the host matrices */
  status = cublasSetVector(dim*dim, sizeof(*psa), psa, 1, pda, 1);
  if (status != CUBLAS_STATUS_SUCCESS) {
      fprintf (stderr, "!!!! device access error (write A)\n");
      return EXIT_FAILURE;
  }

  status = cublasSetVector(dim*dim, sizeof(*pdb), psb, 1, pdb, 1);
  if (status != CUBLAS_STATUS_SUCCESS) {
      fprintf (stderr, "!!!! device access error (write B)\n");
      return EXIT_FAILURE;
  }

  status = cublasSetVector(dim*dim, sizeof(*psc), psc, 1, pdc, 1);
  if (status != CUBLAS_STATUS_SUCCESS) {
      fprintf (stderr, "!!!! device access error (write C)\n");
      return EXIT_FAILURE;
  }

  /* Clear last error */
  cublasGetError();

  /* Performs operation using cublas */
  cublasDgemm('n', 'n', dim, dim, dim, alpha, pda, dim, pdb, dim, beta, pdc, dim);
  status = cublasGetError();
  if (status != CUBLAS_STATUS_SUCCESS) {
      fprintf (stderr, "!!!! kernel execution error.\n");
      return EXIT_FAILURE;
  }

  /* Read the result back */
  status = cublasGetVector(dim*dim, sizeof(*psc), pdc, 1, psc, 1);
  if (status != CUBLAS_STATUS_SUCCESS) {
      fprintf (stderr, "!!!! device access error (read C)\n");
      return EXIT_FAILURE;
  }
  clock_gettime(CLOCK_MONOTONIC, &t2); 
  deltaT = timeDiff(&t1, &t2);
  printf("Done.       Elapsed Time = %6.4f secs\n", deltaT);
  printf("                                  ");
  printf("GFlOP Rate   = %8.4f\n", gflopCnt/deltaT);

  if ( (float) dim - psc[0] > 1.0e-5  ||
       (float) dim - psc[dim*dim-1] > 1.0e-5 ) {
      printf("Error: Incorrect Results!\n");
      printf("C[%2d,%2d] = %10.4f\n", 1,1,psc[0]); 
      printf("C[%2d,%2d] = %10.4f\n", dim,dim,psc[dim*dim-1]); 
  }
}

resuse.c

{{#fileAnchor: resuse.c}} Download raw source of the [{{#fileLink: resuse.c}} resuse.c] 

/*---------------------------------------------------------------------------*
 *                      resuse.c                                             *
 *                                                                           *
 * resuse_(desc)                                                              *
 * char *desc;                                                               *
 *---------------------------------------------------------------------------*
 * This routine makes use of the getrusage system call to collect and        *
 * display resource utilization from one point in a program to another.      *
 * Thus, it works somewhat like a stopwatch in the since that the first call *
 * initializes, or starts, the resource collection over the interval.  The   *
 * second call stops the collection and displays the results accumulated     *
 * since the previous call.  This means, of course, that two calls are       *
 * required for each section of code to be monitored.                        *
 *
 * The underscore at the end of the routine name is there so that the routine*
 * may be called as an integer valued FORTRAN function name RESUSE(), under  *
 * both the SunOS and Ultrix f77 compilers.  AIX inter-language calling      *
 * conventions are different so the routine must be referenced as RESUSE_()  *
 * under AIX (RISC/6000) FORTRAN (xlf).
 *
 * From a FORTRAN routine:
 *
 * INT = RESUSE("Some String")
 * 
 *    fortran code to be timed
 *
 * INT = RESUSE("Some String")
 *
 * Just ignore the returned value.
 *---------------------------------------------------------------------------*/

#include <sys/time.h>
#include <sys/resource.h>

static int first_call = 1;

static struct rusage initial;

void resuse(str)
char *str;
{
  int     pgminor, pgmajor, nswap, nvcsw, nivcsw;
  int     inblock, oublock;
  struct  rusage final;
  float   usr, sys, secs;

  getrusage(RUSAGE_SELF, &final);

  if ( ! first_call )
  {
    secs = final.ru_utime.tv_sec + final.ru_utime.tv_usec / 1000000.0;
    usr  = secs;
    secs = initial.ru_utime.tv_sec + initial.ru_utime.tv_usec / 1000000.0;
    usr  = usr - secs;

    secs = final.ru_stime.tv_sec + final.ru_stime.tv_usec / 1000000.0;
    sys  = secs;
    secs = initial.ru_stime.tv_sec + initial.ru_stime.tv_usec / 1000000.0;
    sys  = sys - secs;

    pgminor = final.ru_minflt  - initial.ru_minflt;
    pgmajor = final.ru_majflt  - initial.ru_majflt;
    nswap   = final.ru_nswap   - initial.ru_nswap;
    inblock = final.ru_inblock - initial.ru_inblock;
    oublock = final.ru_oublock - initial.ru_oublock;
    nvcsw   = final.ru_nvcsw   - initial.ru_nvcsw ;
    nivcsw  = final.ru_nivcsw  - initial.ru_nivcsw ;

    printf("=============================================================\n");
    printf("%s: Resource Usage Data...\n", str);
    printf("-------------------------------------------------------------\n");
    printf("User Time   (secs)          : %10.3f\n", usr);
    printf("System Time (secs)          : %10.3f\n", sys);
    printf("Total Time  (secs)          : %10.3f\n", usr + sys);
    printf("Minor Page Faults           : %10d\n", pgminor);
    printf("Major Page Faults           : %10d\n", pgmajor);
    printf("Swap Count                  : %10d\n", nswap);
    printf("Voluntary Context Switches  : %10d\n", nvcsw);
    printf("Involuntary Context Switches: %10d\n", nivcsw);
    printf("Block Input Operations      : %10d\n", inblock);
    printf("Block Output Operations     : %10d\n", oublock);
    printf("=============================================================\n");

  }
  else
  {
    printf("=============================================================\n");
    printf("%s: Collecting Resource Usage Data\n", str);
    printf("=============================================================\n");
  }

  first_call = !first_call;
  initial = final;

  return;
}

Makefile for Building cuda_bm

{{#fileAnchor: Makefile}} Download raw source of the [{{#fileLink: Makefile}} Makefile] 

#
# Makefile
#
CC          = icc
CFLAGS      = -O3 -openmp 

MKL_INCS    = -I$(HPC_MKL_INC)
MKL_LIBS    = -L$(HPC_MKL_LIB) -lmkl_intel_lp64 -lmkl_sequential -lmkl_core
MKL_MP_LIBS = -L$(HPC_MKL_LIB) -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_core

CUDA_INCS   = -I$(HPC_CUDA_INC) 
CUDA_LIBS   = -L$(HPC_CUDA_LIB) -lcublas 

LIBS        = $(MKL_MP_LIBS) $(CUDA_LIBS) -lpthread -lrt
INCS        = $(MKL_INCS) $(CUDA_INCS)

bm: bm.o ptime.o resuse.o
	$(CC) $(CFLAGS)  -o bm bm.o ptime.o resuse.o $(LIBS)

cuda_bm: cuda_bm.c resuse.o
	$(CC) $(CFLAGS) $(INCS) -o cuda_bm cuda_bm.c resuse.o $(LIBS)

cuda_sp_bm: cuda_sp_bm.c resuse.o
	$(CC) $(CFLAGS) $(INCS) -o cuda_sp_bm cuda_sp_bm.c resuse.o $(LIBS)

clean:
	rm -f *.o core.* bm cuda_bm

Compiling and Linking

Download the above three files into a single directory. Then type: 

[taylor@c11a-s15 Bench]$ module load intel/2013
[taylor@c11a-s15 Bench]$ module load cuda/5.5
[taylor@c11a-s15 Bench]$ make clean
rm -f *.o core.* bm cuda_bm

[taylor@c11a-s15 Bench]$ make cuda_bm
icc -O3 -openmp -I/opt/intel/composer_xe_2013_sp1.0.080/mkl/include   -c -o resuse.o resuse.c
icc -O3 -openmp -I/opt/intel/composer_xe_2013_sp1.0.080/mkl/include -I/opt/cuda/5.5/include -I/opt/intel/composer_xe_2013_sp1.0.080/mkl/include -o cuda_bm cuda_bm.c resuse.o -L/opt/cuda/5.5/lib64 -lcublas -L/opt/intel/composer_xe_2013_sp1.0.080/mkl/lib/intel64 -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_core -lpthread -lrt

Output

# Host: Single, Hex-Core Intel 5675 @ 3.0 GHz
# GPU: Kepler K20c

$ export OMP_NUM_THREADS=4
$ ./cuda_bm
Initializing Matrices...Done.    Elapsed Time = 2.3768 secs
Starting parallel DGEMM...Done.  Elapsed Time = 33.2302 secs
                                 GFlOP Rate   =  45.3546
Initializing CUDA...Done.
Re-initializing Matrices...Done.  Elapsed Time = 2.1733 secs
Starting CUDA DGEMM...Done.       Elapsed Time = 2.3522 secs
                                  GFlOP Rate   = 640.7458



Retrieved from "https://help.rc.ufl.edu/mediawiki/index.php?title=Nvidia_CUDA_Toolkit&oldid=11710"