[Pvfs2-developers] i/o benchmark for pvfs

[Robert Latham]

On Tue, Apr 10, 2007 at 11:33:34AM -0400, vijayaku at cse.psu.edu wrote:
> Hi,
> What would be the best i/o benchmark that one can use to evaluate parallel
> filesystem like pvfs (what has been used until now). Does using Dbench
> make sense? i need more i/o benchmarks.

PVFS comes with a couple simple benchmarks:
- test/client/mpi-io/mpi-io-test.c measures aggregate bandwidth for
  large contiguous parallel accesses via MPI-IO
- test/client/mpi-io/mpi-md-test.c measures MPI-IO metadata operations

LLNL has written a variety of IO benchmarks (IOR is perhaps the most
widely known, but the other tests are good too)
http://www.llnl.gov/icc/lc/siop/downloads/download.html

b_eff_io produces a lot of information, but it's a good one too:
http://www.hlrs.de/organization/par/services/models/mpi/b_eff_io/

dbench doesn't make sense for a parallel file system because it is a
serial application and also focuses heavily on metadata operations.
We do use dbench as a correctness test.

Likewise, bonnie++ and iozone, while very popular benchmarks, are
better suited for local file systems.  See
http://www.pvfs.org/pvfs2-faq.html#sec:badperf for more information.

If you can tell us what sort of workload you want to use against PVFS,
we can let you know some appropriate benchmarks.

==rob

-- 
Rob Latham
Mathematics and Computer Science Division    A215 0178 EA2D B059 8CDF
Argonne National Lab, IL USA                 B29D F333 664A 4280 315B