GPU Guide

From Storrs HPC Wiki
Jump to: navigation, search

Our GPU nodes are available to all users. They are separated into four different partitions: gpu, gpu_v100, gpu_gtx, and gpu_rtx. Each partition has the type of GPU cards shown here.

Compute jobs submitted to these partitions can run for up to twelve hours.

Basic Information

For most work on the GPU nodes, you'll first need to load a CUDA module.

$ module load cuda

This will load the default CUDA module cuda/10.2-rhel7.7.

You can obtain the full list of avilable modules as follows:

$ module avail cuda

To compile with CUDA using the NVidia CUDA compiler:

$ nvcc {} -o {OUTPUT_FILE}

GPU V100 Nodes

A list of our GPU nodes is here.

However, our V100 nodes come in a few varieties. Here are number of V100 cards and the memory on each card for our V100 nodes.

Configuration of each V100 GPU Node
NodeName Number of GPU Memory On GPU
gpu03 3 16GB
gpu04 3 16GB
gpu05 3 16GB
gpu06 1 16GB
gpu07 1 16GB
gpu08 1 16GB
gpu09 3 32GB
gpu10 3 32GB
gpu11 1 16GB

GPU Debug Partition

We also have a GPU partition for debugging GPU jobs, gpu_debug. This partition is available to everyone and has a high priority (though not quite as high as the standard Priority partitions) so your debug jobs can start faster. However, these jobs are limited to one node and a 30 minute run time.

This debug partition applies to all four of our GPU types, so you'll need to use the following --exclude= parameter in your SBATCH file to direct your job to the correct GPU family. See the next section, Batch Jobs, for how to use these parameters.

GPU Type Exclude Line
K40m --exclude=gpu[03-11],gtx[01-16]
V100 --exclude=gpu[01-02],gtx[01-16]
GTX --exclude=gpu[01-11],gtx[12-16]
RTX --exclude=gpu[01-11],gtx[01-11]

Batch Jobs

The script below serves as an example for submitting a job to the scheduler to use a single GPU card, for up to four hours. You should change the variables in bold to meet your needs.

#SBATCH --gres=gpu:NUM_GPUS
#SBATCH --time=04:00:00

Select the value for GPU_PARTITION from this table, or use gpu_debug.

Set NUM_GPUS to the number of GPUs that you want to use, according to the table linked above.

Also, if you are using the gpu_debug partition, you might want to include one of the '--exclude' commands shown in the previous section.

In this example, we set the job's time limit to 4 hours (04:00:00).

Note that for gpu_debug, any time limit greater than 30 minutes (00:30:00) will cause your job submission to fail, since that partition is limited to 30 minutes.

Then, submit the script to the job scheduler:


Interactive Jobs

Assign one of the GPU nodes using fisbatch:

module load cuda/8.0
fisbatch --partition=gpu -c <numprocs> --gres=gpu:NUM_GPUS

Where NUM_GPUS needs to be replaced by the number of CPU processors you need, and NUM_GPUS needs to be replaced by the number of GPUs you need.

Then run the deviceQuery sample program to get useful information about the GPU information like memory, processor cores, etc:

$ /apps2/cuda/8.0/samples/1_Utilities/deviceQuery/deviceQuery
/apps2/cuda/8.0/samples/1_Utilities/deviceQuery/deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

Detected 1 CUDA Capable device(s)

Device 0: "Tesla K40m"
  CUDA Driver Version / Runtime Version          8.0 / 8.0
  CUDA Capability Major/Minor version number:    3.5
  Total amount of global memory:                 11440 MBytes (11995578368 bytes)
  (15) Multiprocessors, (192) CUDA Cores/MP:     2880 CUDA Cores
  GPU Max Clock rate:                            745 MHz (0.75 GHz)
  Memory Clock rate:                             3004 Mhz
  Memory Bus Width:                              384-bit
  L2 Cache Size:                                 1572864 bytes
  Maximum Texture Dimension Size (x,y,z)         1D=(65536), 2D=(65536, 65536), 3D=(4096, 4096, 4096)
  Maximum Layered 1D Texture Size, (num) layers  1D=(16384), 2048 layers
  Maximum Layered 2D Texture Size, (num) layers  2D=(16384, 16384), 2048 layers
  Total amount of constant memory:               65536 bytes
  Total amount of shared memory per block:       49152 bytes
  Total number of registers available per block: 65536
  Warp size:                                     32
  Maximum number of threads per multiprocessor:  2048
  Maximum number of threads per block:           1024
  Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
  Max dimension size of a grid size    (x,y,z): (2147483647, 65535, 65535)
  Maximum memory pitch:                          2147483647 bytes
  Texture alignment:                             512 bytes
  Concurrent copy and kernel execution:          Yes with 2 copy engine(s)
  Run time limit on kernels:                     No
  Integrated GPU sharing Host Memory:            No
  Support host page-locked memory mapping:       Yes
  Alignment requirement for Surfaces:            Yes
  Device has ECC support:                        Enabled
  Device supports Unified Addressing (UVA):      Yes
  Device PCI Domain ID / Bus ID / location ID:   0 / 3 / 0
  Compute Mode:
     < Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 8.0, CUDA Runtime Version = 8.0, NumDevs = 1, Device0 = Tesla K40m
Result = PASS

Then, exist your interactive session.

$ exit
[screen is terminating]
Connection to gpu01 closed.
FISBATCH -- exiting job