[Lustre-discuss] seek heavy workloads and client page cache

[Andreas Dilger]

On Jun 25, 2008  14:30 +0100, Guy Coates wrote:
> I have a seek heavy workload on a couple of files which are small
> enough to fit into the client page cache. If I run the workload on
> local disk, the first run is comparatively slow but subsequent runs
> are much faster as the files are all cached.

> strace snippet of the workload:
> read(4, "z\0\0\0\0\0\0\1\217Q1AWM4.1\0\0\0\0\0\0\0S6\254\20\0\0"..., 2097152) =
> 2097152

The problem is that your application is using the st_blocksize reported
from Lustre (2MB) as the IO size for the application, and this will hurt
if you are continuously reading that much data from even the kernel cache.

The application should really only read as much data as it needs.  It
is possible to tune this by specifying a small stripe size (min 64kB),
which will translate into an small st_blocksize for the file, but it
still isn't as good as fixing the application.

Cheers, Andreas
--
Andreas Dilger
Sr. Staff Engineer, Lustre Group
Sun Microsystems of Canada, Inc.