How to Start a Spark Cluster on Kestrel

This section assumes that you have cloned the HPC Spark setup scripts from this repo. If you are unfamiliar with that, please read the full details at HPC.

Note

The latest scripts currently supporting Kestrel are at this branch: https://github.com/daniel-thom/HPC/tree/kestrel-update

Compute Node Types

Spark works best with fast local storage. The standard Kestrel nodes do not have any local storage. The best candidates are the 256 standard nodes (no GPUs) with 1.92 TB NVMe M.2 drives. Please refer to the Kestrel system configuration page for specific hardware information. The GPU nodes will work as well, but at a greater cost in AUs.

If those nodes are not available, you may be able to complete your queries by using the standard nodes and specifying a path on the Lustre filesystem in the Spark configuration file conf/spark-env.sh. Change SPARK_LOCAL_DIRS and SPARK_WORKER_DIR.

Steps

  1. From the HPC login node, create a work directory somewhere in /scratch/$USER.

  2. Add the spark_scripts to your PATH if you haven’t already:

$ export PATH=$PATH:~/repos/HPC/applications/spark/spark_scripts

  1. Allocate one or more nodes.

$ salloc -t 01:00:00 -N1 --account=dsgrid --partition=debug --tmp=1600G --mem=240G

  1. Configure the Spark settings and start the cluster. Run -h to see the available options.

$ configure_and_start_spark.sh

  1. Set the Spark configuration environment variable.

$ export SPARK_CONF_DIR=$(pwd)/conf

The Spark cluster is ready to query at spark://$(hostname):7077. Run all query scripts from this node.