Difference between revisions of "Getting started with GPUs"

From ScientificComputing
Jump to: navigation, search
(Available GPU node types)
 
(10 intermediate revisions by the same user not shown)
Line 1: Line 1:
 
==Introduction==
 
==Introduction==
There are GPU nodes in both, the Euler and the Leonhard Open cluster. The GPU nodes on both cluster 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.  
+
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==
 
==How to submit a GPU job==
All GPUs are configured in Exclusive Process mode. To run multi-node job, you will need to request <tt>span[ptile=XX]</tt> with <tt>XX</tt> 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).
+
All GPUs in Slurm are configured in non-exclusive process mode. For single node jobs, you can request a number of GPUs with the option <tt>--gpus=''number of GPUs''</tt>
 +
 
 +
sbatch --gpus=''number of GPUs'' ...
  
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 multi-node jobs, you can use the option <tt>--gpus-per-node=''number of GPUs''</tt>
  
For example, to run a serial job with one GPU,
+
  sbatch --gpus-per-node=''number of GPUs'' ...
  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&nbsp;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 [https://devblogs.nvidia.com/parallelforall/cuda-pro-tip-control-gpu-visibility-cuda_visible_devices/ CUDA_VISIBLE_DEVICES] environment variable, which is honored by CUDA programs.
+
or for example in a jobscript
 +
 
 +
#!/bin/bash
 +
 +
#SBATCH --ntasks=8
 +
#SBATCH --nodes=2
 +
#SBATCH --gpus-per-node=1
 +
 +
''command [argument]''
 +
 
 +
This would request 2 nodes, each with 1 GPU and 4 CPU cores.
 +
 
 +
==Sofware with GPU support==
 +
On Euler, packages with GPU support are only available in the [[Euler_applications_and_libraries|new software stack]]. None of the packages in the old software stack on Euler has support for GPUs.
  
 
==Available GPU node types==
 
==Available GPU node types==
 
===Euler===
 
===Euler===
{| class="wikitable"
+
{{GPUTable}}
|-
 
! GPU Model !! Specifier !! GPU memory per GPU !! CPU cores per node !! CPU memory per node
 
|-
 
| NVIDIA RTX 2080 Ti || <tt>GeForceRTX2080Ti</tt> || 11&nbsp;GiB || 36 || 384&nbsp;GiB
 
|-
 
| NVIDIA RTX 2080 Ti || <tt>GeForceRTX2080Ti</tt> || 11&nbsp;GiB || 128 || 512&nbsp;GiB
 
|-
 
| NVIDIA Titan RTX || <tt>TITANRTX</tt> || 24&nbsp; GiB || 128 || 512&nbsp;GiB
 
|-
 
| NVIDIA Tesla V100 || <tt>TeslaV100_SXM2_32GB</tt> || 32&nbsp;GiB || 48 || 768&nbsp;GiB
 
|-
 
| NVIDIA Tesla A100 || <tt> A100_PCIE_40GB </tt> || 40&nbsp;GiB || 48 || 768&nbsp;GiB
 
|}
 
 
 
===Leonhard Open===
 
{| class="wikitable"
 
|-
 
! GPU Model !! Specifier !! GPU memory per GPU !! CPU cores per node !! CPU memory per node
 
|-
 
| NVIDIA GTX 1080 || <tt>GeForceGTX1080</tt> || 8&nbsp;GiB || 20 || 256&nbsp;GiB
 
|-
 
| NVIDIA GTX 1080 Ti || <tt>GeForceGTX1080Ti</tt> || 11&nbsp;GiB || 20 || 256&nbsp;GiB
 
|-
 
| NVIDIA RTX 2080 Ti || <tt>GeForceRTX2080Ti</tt> || 11&nbsp;GiB || 36 || 384&nbsp;GiB
 
|-
 
| [[Nvidia_DGX-1_with_Tensor_Cores| NVIDIA Tesla V100]] || <tt>TeslaV100_SXM2_32GB</tt> || 32&nbsp;GiB || 40 || 512&nbsp;GiB
 
|}
 
  
 
== How to select GPU memory ==
 
== How to select GPU memory ==
 +
If you know that you will need more memory on a GPU than some models provide, <em>i.e.,</em> more than 8&nbsp;GB, then you can request that your job will run only on GPUs that have enough memory. Use the <tt>gpumem:''XX''g</tt> option, where ''XX'' is the amount of GPU memory in GB. For example, if you need 10&nbsp;GB per&nbsp;GPU:
  
If you know that you will need more memory on a GPU than some models provide, <em>i.e.,</em> more than 8&nbsp;GB, then you can request that your job will run only on GPUs that have enough memory. Use the <tt>gpu_mtotal0</tt> host selection to do this. For example, if you need 10&nbsp;GB (=10240&nbsp; MB) per&nbsp;GPU:
+
   [sfux@eu-login-01 ~]$ '''sbatch --gpus=1 --gpumem:10g ./my_cuda_program'''
 
 
   [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&nbsp;GB of GPU memory.
 
This ensures your job will not run on GPUs with less than 10&nbsp;GB of GPU memory.
Line 58: Line 43:
 
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.
 
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 <tt>-R "select[gpu_model1==GPU_MODEL]"</tt> resource requirement to bsub,
+
To select a certain GPU model, use the <tt>--gpus=GPUMODEL:number</tt> resource requirement to bsub,
  
  [sfux@lo-login-01 ~]$ '''bsub -R "rusage[ngpus_excl_p=1]" -R "select[gpu_model0==GeForceGTX1080]" ./my_cuda_program'''
+
  [sfux@eu-login-01 ~]$ '''sbatch --gpus=gtx_1080_ti:1 ./my_cuda_program'''
  
 
==Python and GPUs==
 
==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.
+
We provide separate Python modules (python/XXX and python_gpu/XXX) that point to the same Python installation. The python_gpu modules will in addition automatically load a CUDA, a CUDNN and an NCCL module.
 
 
For an overview on the available Python and TensorFlow versions on Leonhard  Open, please have a look at [[Python_on_Leonhard|Python on Leonhard]]
 
 
 
===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!')
 
<nowiki>matrix1 = tf.constant([[3., 3.]])</nowiki>
 
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.
 

Latest revision as of 07:54, 20 October 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 in Slurm are configured in non-exclusive process mode. For single node jobs, you can request a number of GPUs with the option --gpus=number of GPUs

sbatch --gpus=number of GPUs ...

For multi-node jobs, you can use the option --gpus-per-node=number of GPUs

sbatch --gpus-per-node=number of GPUs ...

or for example in a jobscript

#!/bin/bash

#SBATCH --ntasks=8
#SBATCH --nodes=2
#SBATCH --gpus-per-node=1

command [argument]

This would request 2 nodes, each with 1 GPU and 4 CPU cores.

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

GPU Model LSF Specifier (GPU driver > 450.80.02) Slurm specifier GPU memory per GPU CPU cores per node CPU memory per node
NVIDIA GeForce GTX 1080 NVIDIAGeForceGTX1080 gtx_1080 8 GiB 20 256 GiB
NVIDIA GeForce GTX 1080 Ti NVIDIAGeForceGTX1080Ti gtx_1080_ti 11 GiB 20 256 GiB
NVIDIA GeForce RTX 2080 Ti NVIDIAGeForceRTX2080Ti rtx_2080_ti 11 GiB 36 384 GiB
NVIDIA GeForce RTX 2080 Ti NVIDIAGeForceRTX2080Ti rtx_2080_ti 11 GiB 128 512 GiB
NVIDIA GeForce RTX 3090 NVIDIAGeForceRTX3090 rtx_3090 24 GiB 128 512 GiB
NVIDIA TITAN RTX NVIDIATITANRTX titan_rtx 24 GiB 128 512 GiB
NVIDIA Quadro RTX 6000 QuadroRTX6000 quadro_rtx_6000 24 GiB 128 512 GiB
NVIDIA Tesla V100-SXM2 32 GiB TeslaV100_SXM2_32GB v100 32 GiB 48 768 GiB
NVIDIA Tesla V100-SXM2 32 GB TeslaV100_SXM2_32GB v100 32 GiB 40 512 GiB
Nvidia Tesla A100 (40 GiB) NVIDIAA100_PCIE_40GB a100_40gb 40 GiB 48 768 GiB
Nvidia Tesla A100 (80 GiB) unavailable a100_80gb 80 GiB 48 1024 GiB

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 gpumem:XXg option, where XX is the amount of GPU memory in GB. For example, if you need 10 GB per GPU:

 [sfux@eu-login-01 ~]$ sbatch --gpus=1 --gpumem:10g ./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, use the --gpus=GPUMODEL:number resource requirement to bsub,

[sfux@eu-login-01 ~]$ sbatch --gpus=gtx_1080_ti:1 ./my_cuda_program

Python and GPUs

We provide separate Python modules (python/XXX and python_gpu/XXX) that point to the same Python installation. The python_gpu modules will in addition automatically load a CUDA, a CUDNN and an NCCL module.