HP StorageWorks Scalable File Share - Manual (Page 48)

The Lustre servers were DL380 G5s with two quad-core processors and 16GB of memory, running
RHEL v5.1. These servers were configured in failover pairs using Heartbeat v2. Each server could
see its own storage and that of its failover mate, but mounted only its own storage until failover.
Figure A-2 shows more detail about the storage configuration. The storage comprised a number
of HP MSA2212fc arrays. Each array had a redundant pair of RAID controllers with mirrored
caches supporting failover. Each MSA2212fc had 12 disks in the primary enclosure, and a second
JBOD shelf with 12 more disks daisy-chained using SAS.
Figure A-2 Storage Configuration
Each shelf of 12 disks was configured as a RAID6 vdisk (9+2+spare), presented as a single volume
to Linux, and then as a single OST by Lustre. Each RAID controller of the pair normally served
one of the volumes, except in failover situations.
The FC fabric provided full redundancy at all points in the data path. Each server had two
dual-ported HBAs providing four 4Gb/s FC links. A server had four possible paths to each
volume, which were consolidated using the HP multipath driver based on the Linux device
mapper. We found that the default round-robin load distribution used by the driver did not
provide the best performance, and modified the multipath priority grouping to keep each volume
on a different host FC link, except in failover situations.
Except where noted, all tests reported here used 500GB SATA drives. SATA drives are not the
best performing, but are the most commonly used. SAS drives can improve performance, especially
for I/O workloads that involve lots of disk head movement (for example, small random I/O).
A.2 Single Client Performance
This section describes the performance of the Lustre client. In these tests, a single client node
spreads its load over a number of servers, so throughput is limited by the characteristics of the
client, not the servers.
48 HP SFS G3.0-0 Performance