benchio

Parallel write bandwidth: benchio

This is the benchio parallel write bandwidth benchmark. This evaluates the performance of all MPI processes using collective MPI-IO writes to a single file.

The souce code for the benchio benchmark application can be found at:

• https://github.com/EPCCed/benchio

Benchmark details

The benchmark was compiled with local array size of (256,256,256) on all systems.

On all Lustre systems, the following Lustre settings were used:

• Lustre: 1 MiB stripe size, maximum striping

Analysis

Results are analysed in a Jupyter notebook:

• benchio performance analysis

Parsing of the benchio output file is performed by functions defined in a custom Python module:

• Synthetic benchmark performance analysis Python functions

Compilation details

• ARCHER (EPCC)
• Cirrus (EPCC)
• Athena (HPC Mid+) - No results yet
• CSD3-Skylake (Cambridge)
• Thomas (MMM Hub)

Job scripts

• ARCHER (EPCC)
• Cirrus (EPCC)
• Athena (HPC Mid+) - No results yet
• CSD3-Skylake (Cambridge)
• Thomas (MMM Hub)

Full output data

• ARCHER (EPCC)
• Cirrus (EPCC)
• Athena (HPC Mid+) - No results yet
• CSD3-Skylake (Cambridge)
• Thomas (MMM Hub)