Zero to Hero
Starting from scratch
HPC systems provide value through their ability to run big jobs faster than other systems. It is natural to look for performance metrics; users need to know if your system can run their job, managers want to know if they received good value from their vendor, and so on.
When a system is procured, the vendor often provides peak performance numbers. These are based on the theoretical maximum for each component. Storage benchmarks can often run at 80% or more of the theoretical peak. Good results are almost never achieved on the first attempt. What follows here is a guide to improving benchmark results to get to the desired result.
A normal test of read and write rates begins with a program that creates random data and writes it to a file. The new file is then read by the processor and sent to /dev/null.
The first time you try this, expect to be underwhelmed. It's far from the peak that you expected. What could be wrong? Where is the bottleneck? Is the problem with IO parameters such as the block size, or is there a problem with parallelism?
Start over, restrict your test to one processor and one IO device. Determine the optimal parameters for this configuration. Then see if another processor helps. See if you can double the speed when you are using two IO devices. Find out if one processor can drive two IO devices at twice the speed of one device.
If you have done this sort of thing before, you are probably familiar with the dd command. If not, it’s time to go to the man page for dd and check it out. Then run something like:
dd if=/dev/zero of=/mnt/filesystem/testdir/file bs=4k count=1m
Compare your test program results (if you started with another test program) with dd results. Vary the block size. You should see improvement as you start to add threads, but the returns are probably diminishing. Make sure yournetwork isn’t the bottleneck. If you've been working with networked filesystems before, this may have happened.
Add another client. Take the scripts and configurations you've got from your single client dd script and run on multiple clients. getting somewhere. So you start adding in 4, then 8 then ... clients watching the numbers creep up. The limitations of dd are becoming obvious, it's getting hard to script all this and keep things coordinated.
Running IOR on the Cluster
The first thing to do with IOR is to get it compiled and running. Using the parameters you found best with dd, run IOR and confirm that with 4 or 8 processors, IOR gives results that are very close to those from an ensemble of dd runs.
IOR, IOzone and xdd are the tools that the Hero Run task group uses. Select your favorite and become familiar with it. Write scripts, run them on your cluster, and use them for testing when you get new hardware.
These codes are also useful for diagnostics. For example, if you have one drive group running at half speed and seven drive groups running at full speed, you will have difficulty finding the slow drive group if all of the tests you run are eight-way parallel.
Xyratex Benchmarking Methodology paper