[Linux-cluster] Bottleneck hunting... (GFS Performance)

[Wendy Cheng]

HAWKER, Dan wrote:

>
>The nodes are presently developer boxes, so my users ssh in and compile
>software mostly but also using Samba/NFS to access. Samba/NFS is only
>running on one node (as I believe neither is cluster-aware without
>significant fiddling).
>  
>
You can access NFS over GFS across cluster - that is supported. However 
we currently have an issue with posix lock not able to get propogated 
across multiple GFS servers due to a Linux NLM problem.  There is a 
proposed (NFS-NLM) fix from IBM and University of Michigan that is being 
looked at. Note that this is a linux NFS issue (all linux cluster 
filesystems suffer this bug) - directly accessing GFS servers using 
posix locks *does* work.

Judging by your application, the performance problem you've seen could 
be caused by fragmentation (lots of deleted and new created files). We 
do have a tentative patch that could help with this. If you can captured 
some oprofile outputs when system is in sluggish mode, we may be able to 
identify whether our fix could help.

-- Wendy