[Beowulf] file IO benchmark

Jim Lux James.P.Lux at jpl.nasa.gov
Thu Nov 24 06:11:20 PST 2005


At 01:49 AM 11/24/2005, Toon Knapen wrote:
>Joachim Worringen wrote:
> >
> > Would running some sort of benchmark really help you (or the customer)
> > only because it is well known? I/O performance depends very much on the
> > access style and pattern, and the numbers that a benchmark delivers do
> > not necessarily reflect what your applications really do.
> >
>
>
>Our applications are not _that_ sensitive to I/O performance. We're
>mainly streaming blocks of a matrix temporarily on disk for reading them
>back into memory later on. So pure bandwith is the most important. Which
>bench would you advise?
>
>The problem is that our parallel direct out-of-core solver thus needs to
>store tmp data on disk. We already encountered problems when people are
>using one global NFS mounted filesystem for storing the tmp data of all
>nodes in the parallel run. Of course this results in continuous
>conflicts on the (slow) ethernet network and a disk trying to stream
>multiple big files at the same time. Requesting users to run this
>'standard' benchmark will enable us also to detect these kind of situations.


Why not something fairly crude like measuring "wall clock" time to 
read/write a large file. Do it as a 4 or 5 line shell script.

Or, are there some subtleties in the dynamics that you're trying to capture?

Jim




More information about the Beowulf mailing list