[PVFS2-developers] pvfs2 performance of 0.5.1pre1

neillm at mcs.anl.gov neillm at mcs.anl.gov
Fri Jun 4 11:53:31 EDT 2004


Hey Fan,

On Fri, Jun 04, 2004 at 10:46:00PM +0800, Fan Zhihua wrote:
> I am sorry to show my test results so slowly.

Slowly compared to what?

> 1. PVFS2 Configuration: metasever, dataserver, client are all on one node.
> 1) TroveSyncMode sync
> # time ./create 1000
> real    0m59.199s
> user    0m0.010s
> sys     0m0.010s

<snip>

> 2) TroveSyncMode nosync
> # time ./create 1000
> real    0m15.401s
> user    0m0.000s
> sys     0m0.040s

All this tells me is that nosync is a big win! ;-)

No really, PVFS2 is a *parallel* file system.  Try writing a
(parallel) program that does thousands of creates from multiple vfs
interfaces -- that's where you'll see some benefit.  Some other file
systems are optimized for local and serial performance, so we're not
surprised if they're faster.

For your tests, you should see the best performance with a parallel
program using MPI-IO.  What is the problem you said you're seeing when
using MPI-IO?  [ we should know about it so we can try to fix it! ]


Best regards,
-Neill.


More information about the PVFS2-developers mailing list