Introduction to GPU accelerated jobs

Currently we have 31 nodes in the yoshi cluster (ygpu01-ygpu31) equipped with GPU boards. The exact hardware config is:

• 2x NVidia Tesla M2070
• 2x Xeon X5570
• 24GB RAM
• QDR Infiniband between all GPU nodes

In oder to use the GPU cards, you need to allocate them through the queuing system using the –gres=gpu:2 option. You could also just use one card if you submit with –gres=gpu:1. You also have to explicitly state the partition to run in using –partition=gpu-main (or gpu-test for the GPU test queue).

Here I give a simple example using GROMACS. First I'll use an interactive session to explore the GPU feature, in the end I'll supply a complete batch script for use with sbatch.

<xterm> dreger@yoshi:~/gpu> sinfo | grep gpu gpu-test up 2:00:00 1 idle ygpu01 gpu-main up infinite 30 idle ygpu[02-31] </xterm>

The test partition gpu-test which consists of the single node ygpu01 will most likely be free, since it has a timelimit of 2 hours. So we'll use that for testing:

<xterm> dreger@yoshi:~/gpu> srun –time=02:00:00 –nodes=1 –tasks=8 –gres=gpu:2 –partition=gpu-test –mem=1G –pty /bin/bash dreger@ygpu01:~/gpu> env | grep CUDA CUDA_VISIBLE_DEVICES=0,1 dreger@ygpu01:~/gpu> nvidia-smi Thu Jun 18 14:16:19 2015 +——————————————————+

+——————————-+———————-+———————-+

+——————————-+———————-+———————-+

+—————————————————————————–+

+—————————————————————————–+ </xterm>

The nvidia-smi command gives some information on the GPUs. Currently no process is running on the GPUs. We'll start a simple GROMACS computation:

<xterm> dreger@ygpu01:~/gpu> module load gromacs/non-mpi/4.6.7-cuda dreger@ygpu01:~/gpu> genbox -box 9 9 9 -p -cs spc216 -o waterbox.gro dreger@ygpu01:~/gpu> grompp -f run.mdp -c waterbox.gro -p topol.top dreger@ygpu01:~/gpu> mdrun […] Using 2 MPI threads Using 4 OpenMP threads per tMPI thread

2 GPUs detected:

#0: NVIDIA Tesla M2070, compute cap.: 2.0, ECC: yes, stat: compatible
#1: NVIDIA Tesla M2070, compute cap.: 2.0, ECC: yes, stat: compatible

2 GPUs auto-selected for this run. Mapping of GPUs to the 2 PP ranks in this node: #0, #1 […]

             Core t (s)   Wall t (s)        (%)
     Time:      262.880       34.401      764.2
               (ns/day)    (hour/ns)

Performance: 25.121 0.955 </xterm>

While your jobs run you can log in to the node and call nvidia-smi to see if the GPUs are used at all:

<xterm> dreger@ygpu01:~> nvidia-smi Thu Jun 18 14:25:21 2015 +——————————————————+

+——————————-+———————-+———————-+

+——————————-+———————-+———————-+

+—————————————————————————–+

+—————————————————————————–+ </xterm>

Please check your job logfiles to see if your program has some problems using the GPUs. In case of GROMACS this might look like:

<xterm> NOTE: GPU(s) found, but the current simulation can not use GPUs To use a GPU, set the mdp option: cutoff-scheme = Verlet (for quick performance testing you can use the -testverlet option)

Using 8 MPI threads

2 GPUs detected:

#0: NVIDIA Tesla M2070, compute cap.: 2.0, ECC: yes, stat: compatible
#1: NVIDIA Tesla M2070, compute cap.: 2.0, ECC: yes, stat: compatible

2 compatible GPUs detected in the system, but none will be used. Consider trying GPU acceleration with the Verlet scheme! </xterm>

In this case a cutoff-scheme was specified that can not be used with GPU acceleration.

Compare the timings with a test run on the same node, that does not use the GPUs. In some cases the GPUs will not help at all, even though nvidia-smi shows a high utilization. For this example without GPU (note the missing -cuda in the module load command) we get:

<xterm> dreger@ygpu01:~/gpu> module load gromacs/non-mpi/4.6.7 dreger@ygpu01:~/gpu> grompp -f run.mdp -c waterbox.gro -p topol.top dreger@ygpu01:~/gpu> mdrun

             Core t (s)   Wall t (s)        (%)
     Time:      844.970      106.315      794.8
               (ns/day)    (hour/ns)

Performance: 8.128 2.953 </xterm>

So in this case the calculation runs about three times faster with two GPU cards.

A job script for the example given above could look like:

<xterm> #!/bin/bash

#SBATCH –mail-user=dreger@physik.fu-berlin.de #SBATCH –mail-type=end

#SBATCH –output=job%j.out #SBATCH –error=job%j.err #SBATCH –ntasks=8 #SBATCH –mem-per-cpu=1024 #SBATCH –time=01:00:00 #SBATCH –gres=gpu:2 #SBATCH –nodes=1 #SBATCH –partition=gpu-main

module load gromacs/non-mpi/4.6.7-cuda

TAG="${SLURM_JOB_ID}-$(hostname -s)-cuda"

grompp -f run.mdp -c waterbox.gro -p topol.top -o output-$TAG mdrun -nt ${SLURM_CPUS_ON_NODE} -testverlet -v -deffnm output-$TAG </xterm>

Please make sure you change the email if you use this for your own tests ;)