I've just brought into service a new SE, se03.esc.qmul.ac.uk at QMUL. It
runs Storm/Lustre
As well as "gsiftp" and "rfio" protocols, storm supports the "file"
protocol.
If user jobs can be persuaded to use the file protocol to access files,
then they will see much better performance. I reached an aggregate
performance of around 3.5GB/s to 60 machines[1] using lustre. I'd have
tried more machines, but ran into bugs in the benchmarking software.
With rfio, or gsiftp, the limit is a 1Gbit network connection providing
around 0.1GB/s.
How do I persuade jobs to use the file protocol? Is it something I must
do, or something the jobs themselves must do?
[1] Approximately - I don't have the exact figures to hand.
|