We are planning to expand our lustre setup. Do anyone have any advice on what
will be the most efficient solution for OSS server hardware?
Should we buy few fat nodes, for ex. with two quad core cpus with lots of
memory? Or rather many simple boxes with less cpus, one single core or dual
core cpu and little memory (1 GB per node).
As it looks now we can either buy two fat OSSs with two dual FC controllers
each to the raids. Or, four simple pizza boxes with one dual FC controller
each to the raids.
Using 1.5TB per OST we get a total of 8 OSTs in this system, so will fewer
OSTs per OSS be beneficiary?
Four simple boxes seems to be cheaper than two fat nodes.
Bugzilla entry 2573 indicates that large memory pools are of little benefit on
the OSS side. Do anyone have a rule of thumb for the recommended memory size
per OST on an OSS?
We use infiniband between the clients and the OSS/MDS.
Any insight is greatly appreciated.
We do this extension primarily to resolve an issue with our current lustre
setup where we run two file systems, home and scratch, on a shared dual
server setup where the MDS and OSS are on the same physical hardware. When
we have high io on scratch the home file system becomes very inresponsive for
interactive use, ls on a catalog with 200 files takes 60 seconds and so on.
We hope to solve this by moving the scratch OSSs out of this hardware and
only keep the scratch MDS on the old system.
Regards,
r.
--
The Computer Center, University of Troms?, N-9037 TROMS? Norway.
phone:+47 77 64 41 07, fax:+47 77 64 41 00
Roy Dragseth, High Performance Computing System Administrator
Direct call: +47 77 64 62 56. email: royd@cc.uit.no