Skip to content

Latest commit

 

History

History
44 lines (23 loc) · 2.21 KB

README.md

File metadata and controls

44 lines (23 loc) · 2.21 KB
Raw

In this benchmark, we deploy containerized GPU applications on a cluster node, On the same node, we're going to submit LULESH and MILC as a traditional batch job.

This benchmark has been tested and evaluated on Piz Daint system with SLURM and node sharing enabled.

Preparation

  1. Build extended rFaaS with remote memory functions by running build.sh in src/remote_memory. To configure rFaaS dependencies, please inspect the README of the package. It should be sufficient to set PKG_CONFIG_PATH environment variable to point to installation directories of pistache, rdmacm, and ibverbs.

  2. Build LULESH from sources in external/LULESH (build scripts are provided).

  3. Build MILC from sources in external/MILC (build scripts are provided).

Preparing baseline

  1. SSH to the node where you're going to deploy the co-located batch job and remote memory function.

  2. Verify that Docker is running correctly on the system.

  3. Create a copy of devices.json, and change in it the IP address of the network interface used for communication.

  4. Start the rFaaS executor with <rfaas-build-dir>/bin/executor_manager --config exec_manager.json --device-database devices_copy.json --skip-resource-manager.

Preparing the client.

  1. SSH to the second node.

  2. Create a copy of devices.json, and change in it the IP address of the network interface used for communication.

  3. In the exec_db.json, change the executor address to the address of rFaaS executor obtained in the previous part.

  4. Start the rFaaS benchmark with <rfaas-build-dir>/benchmarks/rma --config benchmark.json --device-database devices_copy.json --name empty --functions examples/libfunctions2.so --executors-database exec_db.json -s 1024 --read_size <size> --rdma_type <rdma-type> --pause <pause> --rma_address <executor-ip-address>

  5. To repeat results from the paper, change <size>, <pause>, and <rdma-type> to use read and write operations of different size, launched with different pause between consecutive invocations.

Launch the batch job

  1. In slurm_lulesh_27.sh, change nodelist to use the node hosting co-located rFaaS executor and LULESH.

  2. Launch the co-located LULESH sbatch < slurm_lulesh_27.sh.

  3. Stop the client benchmark.