On Jan 13, 2010, at 4:14 PM, Milo wrote:

> Hi, guys. We're doing research enabling HPC applications on cloud 
> filesystems. For comparison purposes, we're using a PVFS2 filesystem as an 
> example of a mature HPC filesystem. However, we're getting unusually bad 
> performance from our cluster on the MPI applications we're testing writing to 
> PVFS2 through mpich2, and want to know if there are any configuration 
> parameters we might be forgetting to tweak.
> 
> We've tried two applications: LANL's mpi_io_test 
> (http://institutes.lanl.gov/data/software/src/mpi-io/README_20.pdf), and the 
> IOR benchmark. Both are running through the mpich2 library, compiled with 
> PVFS2 support. In other words, we're connecting to pvfs2 with mpi rather than 
> through a kernel module. We're using version 2.8.1 of PVFS2, configured with 
> a stripe size of 1 MB. The mpi_io_test program is writing 1 MB objects.

Are you using collective or independent I/O with IOR?

> 
> The cluster has a 1gig E network. We vary the number of servers/clients from 
> 1 to 40. Every server handles both metadata and data. Clients and servers 
> share the same nodes (this makes things more comparable to the typical cloud 
> computing situation).
> 
> The performance we're seeing is on the order of 15 to 20 MB/s/node. This is 
> much lower than the performance we're getting through network tools such as 
> crisscross or iperf (which report 45-50 MB/s between most pairs of 
> nodes--still too low but more than 20 MB/s).

You could try bmi_pingpong -- it would give you a better idea of what to expect 
out of BMI, and then maybe pvfs2-cp of a file with 1MB requests.  That would 
give you a better idea of what to expect out of the MPI-IO layer.  

> 
> It doesn't seem to be a disk speed issue: We've tested with the TroveMethod 
> set to null-aio and metadata syncing turned off and still get the same 
> results. So we believe it might be a networking issue with the way we're 
> sending MPI messages. Perhaps they're too small?

It would be strange if MPI messages were slowing you down.  IOR won't send MPI 
messages unless collective I/O is used.

-sam

> 
> Any advice as to parameters we should look into tweaking for MPI to get 
> performance closer to that of the network? Any help is appreciated. Thanks.
> 
> ~Milo and Esteban
> _______________________________________________
> Pvfs2-users mailing list
> [email protected]
> http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users


_______________________________________________
Pvfs2-users mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users

Reply via email to