slurm-FFTW-perf-tester

This test measures the FFTWs performance on Linux Clusters using SLURM

It was run on the Leibniz Rechenzentrum Clusters in Germany, but should also work on other clusters supporting FFTW+MPI

What does this test do?

This test was created to simulate the behavior of a fiber-simulation tool at my university, so it does rather specific stuff, but might be useful for others as well.

Basically, it does:

• Allocate an amount of memory corresponding to a power of 2
• Set each byte to 0x00
• Use these bytes to calculate the FFTWs and reverse FFTWs for a certain number of times
• Measure the total program runtime.
• Log the time into logfiles. Wohoo.

Configurating

The Config File

For each node-amount-folder-thingy there's a separate config file.

The config file contains three parameters:

1. The exponent for the power of 2 (amount of RAM)
2. The number of threads to use per node
3. How often the FFTW-invFFTW-pairs should be done

The order of these parameters is important, because I was too lazy to write a proper parser.

The SLURM Script

Note: Don't forget to add your e-mail-adress in your scripts! Otherwise the cluster won't inform you that your job has failed and you will be too proud of yourself.

There's also a slurm script for each node configuration in the subfolder. You are responsible for adjusting:

• the runtime
• the minimum number of threads so
• the approx. needed amount of RAM

The number of threads must always be at least as large as the number specified in the config file.

Building

For LRZ systems:

module load fftw/3.3.8-intel-impi
make

Running

• If you run for the first time: make prepare
• switch to the directory indicating how many nodes you want to use
• edit the config file
• edit the slurm script (.cmd)
• Run it: sbatch SLURMSKRIPT.cmd