Running the Applications#

Given the number of applications, options, datasets, and GPU configurations, we have tried to simplify application testing as much as possible. To facilitate test sweeps across 1 to 16 GPUs, multiple application options, and datasets, every application has two associated scripts: hive-mgpu-run.sh and hive-application-test.sh.

In general hive-mgpu-run.sh deals with parameter sweeps and schedules hive-application-test.sh as multi-GPU SLURM jobs. The hive-application-test.sh script generally deals with datasets and associated paths, and configures itself with the parameters necessary to run the application.

Default Run Configuration#

The simplest way to run an application is to execute:

./hive-mgpu-run.sh

The associated hive-application-test.sh will execute with datasets in the following user directories on NVIDIA’s nslb cluster:

/home/u00u7u37rw7AjJoA4e357/data/gunrock/gunrock_dataset
/home/u00u7u37rw7AjJoA4e357/data/gunrock/hive_datasets

Alternate Run Configurations#

Additional command line parameters and / or script modifications are necessary to run on additional datasets or with alternate application parameters.

hive-mgpu-run.sh#

This script configures SLURM with NUM_GPUS to sweep across on a chosen PARTITION_NAME. Running the script with no parameters (as shown above) is equivalent to:

./hive-mgpu-run.sh 16 dgx2

This runs hive-application-test.sh across 1 to 16 GPUs on the machine partition named dgx2.

For some applications, this script might have additional parameter variables that are worth exploring and modifying. Please see the individual HIVE application chapters for more details.

hive-application-test.sh#

The primary reason to modify this script is to provide additional dataset information. In general these scripts will include some or all of the following arrays:

  • DATA_PREFIX path to directory containing desired dataset

  • NAME a simple string naming the dataset, generally sans a file extension (e.g., NAME[0]="twitter" for twitter.mtx)

  • GRAPH aggregated options for the chosen dataset to pass to the application (i.e., combine DATA_PREFIX and NAME with additional information expected by the application)

Please note that you must update the associated for loop index if you add or remove items to the arrays mentioned.

Future Script Simplification#

In the future we would like to refactor hive-mgpu-run.sh to simply configure the necessary SLURM command (e.g., resources and hardware partition) and pass the command to the hive-application-test.sh script. The application script can then deal with sweeping across its relevant parameters and datasets.