Lustre discuss - Jun 2006 - Client & OST on same node (Revisited)

Hello.

Wanted to do a quick revisit to the subject of running the client and
OST on the same node. 

I see that both the stable and beta download pages say "Running a client
and object server on the same node is known not to be 100% stable;
application or system hangs are possible."

Could someone enlighten me on the issues related to doing so?

The reason to do this is that we would like to be able to use Lustre on
mobile demo-clusters.  These clusters typically have 6-8 nodes (dual
core, dual drives, 4-8GB of memory) which is perfect for the application
but leaves little room for extra storage nodes.

Your input, pls.

Thanks for taking the time!

Best regards,
Diderich Buch

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://mail.clusterfs.com/pipermail/lustre-discuss/attachments/20060626/771ef3eb/attachment.html

On Jun 26, 2006  08:07 +0200, DB wrote:
> I see that both the stable and beta download pages say "Running a
client
> and object server on the same node is known not to be 100% stable;
> application or system hangs are possible."
> 
> Could someone enlighten me on the issues related to doing so?
Consider case of "client" with dirty filesystem pages in memory and
memory
pressure.  A kernel thread is woken to flush dirty pages to the filesystem,
and it writes to local OST.  The OST needs to do an allocation in order to
complete the write.  The allocation is blocked waiting for the above kernel
thread to complete the write and free up some memory.  This is a deadlock.
> The reason to do this is that we would like to be able to use Lustre on
> mobile demo-clusters.  These clusters typically have 6-8 nodes (dual
> core, dual drives, 4-8GB of memory) which is perfect for the application
> but leaves little room for extra storage nodes.
This will generally work (in fact a lot of basic testing is done this way).
There is just a potential deadlock situation, and it becomes more likely
the more heavy IO that is done on the client.

Cheers, Andreas
--
Andreas Dilger
Principal Software Engineer
Cluster File Systems, Inc.