Different benchmarks that can be used on the machines at FI. Note that it is easy to overwhelm the batch schedulers, so try and make sure to not run something that will launch hundreds of jobs. Using disBatch is a good idea if you can (single node)
# JUBE: These are different benchmarks to be used with the JUBE software [Download JUBE here](https://www.fz-juelich.de/ias/jsc/EN/Expertise/Support/Software/JUBE/_node.html)The idea is to provide an input file describing what you want to test (eg: different inputs, number of nodes, compilers, libraries...) and JUBE will run a matrix of all the different possible combinations, then present the results in a structured way.
A typical run will look like this:
jube run definition.xml # Launch the benchmark
jube definition npb_omp --id 0 # Check periodically the progress of the benchmark
jube result definition --id 0 # Get the results
These folders do not contain any packages: you might need to download them if they are not present in the modules system. Input files, XML files for the benchmarks, and Slurm templates are provided.
The main goal is to show different ways of using JUBE, to check performance and scaling, compiler settings, environment settings, etc
This test is of interest for anyone who wants to compare the performance of the code generated by different compilers. The benchmark will start by compiling the code, then run on different problem sizes, increasing with the number of processors
This test can be used for scalability testing, testing different ranks/threads configurations, on both GPU and CPU based clusters, for different inputs
This test contains both the (non optimized) reference code for HPCG, but also the NVIDIA provided one for GPUs. This shows how to use singularity.
This test contains also the reference code, as well as the NVIDIA optimized version which has to be run through a singularity
Different versions of the NPB:
- single-node disBatch: when running with up to number of cores per node, we can run our suite using disBatch, this is such an example. Submit with mpi_singlenode_disbatched.slurm instead of jube directly
- Generic MPI and OpenMP:
- Single-node
- Multi-node (MPI only)
Several sample template submission files are provided (to be placed in ~/.local/lib/python3.7/site-packages/mdbenchmark/templates/):
- rusty2021.1_rome to test performance on the rome nodes and assess the best ratio of MPI ranks vs OpenMP threads
- rusty2021.1_gpu can be used to test the performance on the GPU nodes
- rusty2021.1_singlegpunode.a100 is to be used on a single node, with the NVIDIA optimizations for Gromacs 2021+