Performance results of LeanMD on the Joliot-Curie supercomputer using different load balancing algorithms

This dataset contains the raw output files generated from the execution of LeanMD on the Joliot-Curie supercomputer (20 SKL Irene nodes), the scripts used to generate them, and the scripts used to parse these results for statistical analysis and plotting.

Software information:

• OS: Red Hat Enterprise Linux 7.6
• OpenMPI: version 2.0.4
• Compilers: C/C++ Intel 17.0.6.256
• Charm++ version: v6.9.0-rc3, build mpi-linux-x86_64 --with-production
• LeanMD source: https://charm.cs.illinois.edu/gerrit/gitweb?p=benchmarks/leanmd.git
• Additional load balancers source: https://github.com/viniciusmctf/packing-schemes/tree/packs_2019-v1
• Charm++, LeanMD, and the load balancers were compiled with -O3

File information:

The raw result files are organized in four directories (oct18, oct23, oct24, and oct24_2).
Each directory contains the results of one batched execution in the supercomputer.
Each batch is composed of 10 repetitions of a set of experiments.
Each set of experiments includes different load balancing algorithms and different problem sizes.
Each set is randomly ordered to avoid interference coming from a specific order of execution.
Each raw file contains the appended output of the application and its load balancer for all 10 repetitions.
The name of the files indicate the load balancer and size of the problem.
For instance, `PackStealLB.240` means that the application was run with PackStealLB and the problem size parameter is 240.

Problem sizes:

• 80: 80×11×5 cells of dimensions 15×15×30
• 120: 120×11×5 cells of dimensions 15×15×30
• 160: 160×11×5 cells of dimensions 15×15×30
• 240: 240×11×5 cells of dimensions 15×15×30
• 320: 320×11×5 cells of dimensions 15×15×30

Each execution of LeanMD ran for 301 iterations with load balancing calls at iterations 40, 140, and 240.

Raw output files:

Each raw output file starts with a Charm++ header providing information on the execution.

For each step of the application, its execution time in ms is provided.

Load balancing calls usually provide information about their start time, end time, and duration. Depending on the load balancer, more information is provided.

Each raw file contains ten executions of the application with a given load balancer and input size.

Generating plots:

The analysis of the results can be done by running the Jupyter notebook named "Analysis of load balancing results.ipynb"