Sorry, by PGI, I meant PGFS...an e-mail had come in about PGI while I
was typing, and my brain swapped things....

My interconnect between all nodes is 10Gbps ethernet.

The main reason is fear of downtime (last time I did an upgrade, the
servers went and did their things for a couple hours, and one of them
didn't complete successfully, resulting in way too much downtime).

command was time cp /mnt/pvfs2/testing/10gbtest /scratch/

network bandwidth was monitored in another window with bwm-ng, showed 11-30MB/s

Top was also used for CPU monitoring, and pvfs usage was high.

On another system with some user jobs, I was seeing 8-11MB/s network
throughput (inbound) and pvfs2-client-core using 66% of a core.  I
thought that was problematic.

--Jim

On Thu, Aug 16, 2012 at 10:00 AM, Becky Ligon <[email protected]> wrote:
> Hi Jim!
>
> What is PGI?
>
> What is your interconnect between compute nodes and I/O servers?
>
> Is there some reason why you didn't upgrade the I/O servers to 2.8.6? (Just
> curious)
>
> Can you send me the exact commands that you are using for your copy tests?
>
> Are your users able to access the directories that were inaccessible before
> the upgrade?
>
> What is your linux version now that you've upgraded?  Is it still
> 2.6.32-220.13.1.el6?
>
> Are you now able to mount-on-boot?
>
> Can you attach to the pvfs2-client-core process using gdb to see where it
> seems to be spending its time?
>
> Becky
>
> On Thu, Aug 16, 2012 at 12:00 PM, Jim Kusznir <[email protected]> wrote:
>>
>> Hello:
>>
>> I'm still struggling to get performance from PGI that's close to local
>> single-disk I/O speeds.
>>
>> Presently, on my newly-rebuilt cluster with orangefs 2.8.6 clients /
>> 2.8.5 servers (3), I am seeing about 68% CPU core loading from
>> pvfs2-client-core, and about 11MB/s throughput on the network.
>>
>> In other tests, trying to copy a 10gb file from pvfs to local, I maxed
>> out at about 30MB/s, with 20MB/s more typical.
>>
>> I feel I should be seeing significantly better speeds than this...And
>> that pvfs should not be so client-cpu-intensive.  Suggestions as to
>> how to debug?
>>
>> I'm also getting reports from users that they note a significant
>> runtime improvement of their jobs if they switch from using pvfs2 to
>> using NFS home directories...I also think that should be wrong.
>>
>> Thanks!
>> --Jim
>> _______________________________________________
>> Pvfs2-users mailing list
>> [email protected]
>> http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
>
>
>
>
> --
> Becky Ligon
> OrangeFS Support and Development
> Omnibond Systems
> Anderson, South Carolina
>
>
_______________________________________________
Pvfs2-users mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users

Reply via email to