nSTREAM

Written by Agustin Forero for Argonne National Laboratory.

A fork of STREAM meant to benchmark and graph the memory bandwidth of a supercomputer. Utilizes OpenMP to get the most out of your HPC system.

Commands include:

• ./generate.sh (-n) (optional-threads): run the benchmark. Assumes number of threads to use from return of $(nproc); otherwise, you can specify a custom amount of threads. You can use the -n flag to disable graphing.
• ./reset.sh: removes everything in the /runs/ directory.

Some things to know:

• It's best to use as many threads as possible. Use something like $(htop), $(lscpu) or $(lstopo) to gauge how many threads you can use.
• nstream.F90 will use increasingly large memory footprints to benchmark. By default, it runs 125 times, starting at n = 100000 and ending at n = 12500000, where 100000n ≈ 2.272MB. Ideally, this is used to come into contact with all available caches before deferring to RAM.
• If needed, you can edit the begin, step and fin values for the benchmark at the top of generate.sh.
• It's difficult for stream, either the version contained herein or from the original repo, to gauge L1 cache performance.
• ./generate.sh will save the results in a folder named from the PID; you can find these results in /runs/.
• The construction of the graph utilizes matplotlib, which must therefore be installed onto your environment. This can be done with tools like Miniconda. Alternatively, you can forgo graphing entirely with the -n flag.
• This repo could, hypothetically, be used to test a normal machine.

Sample output of ./generate.sh: