Difference between revisions of "Getting started with GPUs"

Revision as of 14:14, 31 May 2022

Introduction

There are GPU nodes in the Euler cluster. The GPU nodes are reserved exclusively to the shareholder groups that invested into them. Guest users and shareholder that purchase CPU nodes but no GPU nodes cannot use the GPU nodes.

CUDA and cuDNN

cuDNN versions provided are compiled for a particular CUDA version. We will soon add here a table with the compatible versions

How to submit a GPU job

All GPUs are configured in Exclusive Process mode. To run multi-node job, you will need to request span[ptile=XX] with XX being the number of CPU cores per GPU node, which is depending on the node type (the node types are listed in the table below).

The LSF batch system has partial integrated support for GPUs. To use the GPUs for a job node you need to request the ngpus_excl_p resource. It refers to the number of GPUs per node. This is unlike other resources, which are requested per core.

For example, to run a serial job with one GPU,

bsub -R "rusage[ngpus_excl_p=1]" ./my_cuda_program

or on a full node with all 8 GeForce GTX 1080 Ti GPUs and up to 90 GB of RAM,

bsub -n 20 -R "rusage[mem=4500,ngpus_excl_p=8]" -R "select[gpu_model0==GeForceGTX1080Ti]" ./my_cuda_program

or on two full nodes:

bsub -n 40 -R "rusage[mem=4500,ngpus_excl_p=8]" -R "select[gpu_model0==GeForceGTX1080Ti]" -R "span[ptile=20]" ./my_cuda_program

While your jobs will see all GPUs, LSF will set the CUDA_VISIBLE_DEVICES environment variable, which is honored by CUDA programs.

Sofware with GPU support

On Euler, packages with GPU support are only available in the new software stack. None of the packages in the old software stack on Euler has support for GPUs.

Available GPU node types

Euler

How to select GPU memory

If you know that you will need more memory on a GPU than some models provide, i.e., more than 8 GB, then you can request that your job will run only on GPUs that have enough memory. Use the gpu_mtotal0 host selection to do this. For example, if you need 10 GB (=10240  MB) per GPU:

 [sfux@lo-login-01 ~]$ bsub -R "rusage[ngpus_excl_p=1]" -R "select[gpu_mtotal0>=10240]" ./my_cuda_program

This ensures your job will not run on GPUs with less than 10 GB of GPU memory.

How to select a GPU model

In some cases it is desirable or necessary to select the GPU model on which your job runs, for example if you know you code runs much faster on a newer model. However, you should consider that by narrowing down the list of allowable GPUs, your job may need to wait for a longer time.

To select a certain GPU model, add the -R "select[gpu_model1==GPU_MODEL]" resource requirement to bsub,

[sfux@lo-login-01 ~]$ bsub -R "rusage[ngpus_excl_p=1]" -R "select[gpu_model0==GeForceGTX1080]" ./my_cuda_program

Python and GPUs

Because some Python packages need different installations for their CPU and GPU versions, we decided to have separate Python modules (python/XXX and python_gpu/XXX) with regards to using CPUs and GPUs. The python_gpu modules will in addition automatically load a CUDA and a CUDNN module. When running the GPU version of TensorFlow (<2.0.0) or PyTorch on a CPU node will immediately crash, because those packages are checking on start up if the compute node has a GPU driver installed. From TensorFlow 2.0.0 on, google merged the CPU and the GPU version of TensorFlow into a single package, but for PyTorch there are still two installations (CPU/GPU) required.

Tensorflow 1.x example

As an example for running a TensorFlow job on a GPU node, we are printing out the TensorFlow version, the string Hello TensorFlow! and the result of a simple matrix multiplication:

[sfux@lo-login-01 ~]$ cd testrun/python
[sfux@lo-login-01 python]$ module load python_gpu/2.7.13
[sfux@lo-login-01 python]$ cat tftest1.py
#/usr/bin/env python
from __future__ import print_function
import tensorflow as tf

vers = tf.__version__
print(vers)
hello = tf.constant('Hello, TensorFlow!')
matrix1 = tf.constant([[3., 3.]])
matrix2 = tf.constant([[2.],[2.]])
product = tf.matmul(matrix1, matrix2)

sess = tf.Session()
print(sess.run(hello))
print(sess.run(product))
sess.close()
[sfux@lo-login-01 python]$ bsub -n 1 -W 4:00 -R "rusage[mem=2048, ngpus_excl_p=1]" python tftest1.py
Generic job.
Job <10620> is submitted to queue <gpu.4h>.
[sfux@lo-login-01 python]$ bjobs
JOBID      USER      STAT  QUEUE      FROM_HOST   EXEC_HOST   JOB_NAME   SUBMIT_TIME
10620      sfux      PEND  gpu.4h     lo-login-01             *tftest.py Sep 28 08:02
[sfux@lo-login-01 python]$ bjobs
JOBID      USER      STAT  QUEUE      FROM_HOST   EXEC_HOST   JOB_NAME   SUBMIT_TIME
10620      sfux      RUN   gpu.4h     lo-login-01 lo-gtx-001  *ftest1.py Sep 28 08:03
[sfux@lo-login-01 python]$ bjobs
No unfinished job found
[sfux@lo-login-01 python]$ grep -A3 "Creating TensorFlow device" lsf.o10620
2017-09-28 08:08:43.235886: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1045] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX 1080, pci bus id: 0000:04:00.0)
1.3.0
Hello, TensorFlow!
12.
[sufx@lo-login-01 python]$

Please note, that your job will crash if you are running the GPU version of TensorFlow on a CPU node, because TensorFlow is checking on start up if the compute node has a GPU driver.

Tensorflow 2.x example

Tensorflow 2.x does no longer use sessions. Please find below an updated example job for tensorflow 2.0, where we create two 2000x2000 Matrices with random numbers and then carry out a matrix multiplication once on the CPU and once on the GPU and then compare the run times.

[sfux@lo-login-01 ~]$ cd testrun/tf/test2
[sfux@lo-login-01 test2]$ module load gcc/6.3.0 python_gpu/3.7.4

The following have been reloaded with a version change:
  1) gcc/4.8.5 => gcc/6.3.0

[sfux@lo-login-01 test2]$ cat tf2test.py 
#/usr/bin/env python

import time
import tensorflow as tf

k = 2000
a = tf.random.uniform(shape=[k,k], minval=0, maxval=20,dtype=tf.float16)
b = tf.random.uniform(shape=[k,k], minval=0, maxval=20,dtype=tf.float16)

cpu_slot = 0
gpu_slot = 0

# Using CPU at slot 0
with tf.device('/CPU:' + str(cpu_slot)):
    start = time.time()
    c1 = tf.matmul(a,b)
    print("Time on CPU:")
    end = time.time() - start
    print(end)

# Using the GPU at slot 0
with tf.device('/GPU:' + str(gpu_slot)):
    start = time.time()
    c2 = tf.matmul(a,b)
    print("Time on GPU:")
    end = time.time() - start
    print(end)

[sfux@lo-login-01 test2]$ bsub -n 1 -W 4:00 -R "rusage[mem=2048, ngpus_excl_p=1]" python tf2test.py 
Generic job.
Job <5074756> is submitted to queue <gpu.4h>.
[sfux@lo-login-01 test2]$ bjobs
JOBID      USER    STAT  QUEUE      FROM_HOST   EXEC_HOST   JOB_NAME   SUBMIT_TIME
5074756    sfux    PEND  gpu.4h     lo-login-01             *f2test.py Mar  5 12:28
[sfux@lo-login-01 test2]$ bjobs
JOBID      USER    STAT  QUEUE      FROM_HOST   EXEC_HOST   JOB_NAME   SUBMIT_TIME
5074756    sfux    RUN   gpu.4h     lo-login-01 lo-s4-082   *f2test.py Mar  5 12:28
[sfux@lo-login-01 test2]$ bjobs
No unfinished job found
[sfux@lo-login-01 test2]$ grep -A1 "Time on" lsf.o5074756
Time on CPU:
63.97628474235535
Time on GPU:
0.4504997730255127
[sfux@lo-login-01 test2]$

With TensorFlow 2.0 it is possible to build a single Python package that supports CPU and GPU. If TensorFlow 2.0 is imported on a pure CPU compute node, it will no longer fail due to checking the GPU driver as it will fall back to the CPU version in this case.