[Lustre-discuss] OSS on compute nodes

Brian J. Murrell Brian.Murrell at Sun.COM
Fri Apr 11 09:34:41 PDT 2008


On Fri, 2008-04-11 at 20:29 +0400, Peter Avakian wrote:
> 
> Do you see any problem in having each compute node, within a grid,   
> acting as an OSS server via the separate IB channel on the fabric? My  
> compute nodes have built-in raid controllers.

If by compute nodes you mean Lustre clients, then yes, this is a problem
and an unsupported configuration.  The reason is because memory
pressures on a client/OSS machine can cause a deadlock.

The client tries to flush pages to an OST to relieve memory pressure.
An OST needs to allocate memory in order to process page flushes from a
client.  If a client trying to relieve memory pressure tries to flush
pages to an OST on the same node, the OST will get failures trying to
allocate memory (which is already under pressure) to fulfill the request
from the client.  Deadlock.

b.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: This is a digitally signed message part
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20080411/59425d72/attachment.pgp>


More information about the lustre-discuss mailing list