Re: [HPDD-discuss] Lustre Pools vs Multiple Lustre File System recommendation
On Jul 22, 2014, at 4:28 PM, Marcus Nilsson <marcus.nilsson(a)atea.se>
wrote:
I would also consider splitting SAS and SATA to dedicated OSS
servers. The reason is that the performance might get impacted on both pools even if only
one of them is getting hammered. Especially if the disks are connected to the same RAID
controller.
That is a good point. On our Lustre 1.8 file system, I have seen cases where hundreds of
clients issue read requests to a single OST. When this is happening, all the OSS service
threads get tied up serving requests to that one OST and the throughput to other OSTs on
the same host is very low. (I don't know if the Lustre 2.x behavior is much
different.)
If this becomes a problem, it might be worthwhile looking into the Network Request
Scheduler (NRS) feature in Lustre 2.4. This could be used to apply a policy that helps
mitigate that kind of uneven behavior.
--
Rick Mohr
Senior HPC System Administrator
National Institute for Computational Sciences
http://www.nics.tennessee.edu