HP StorageWorks Scalable File Share - Manual (Page 65)

Figure A-3 Single Stream Throughput
For a file written on a single OST (a single RAID volume), throughput is in the neighborhood of
200 MB/s. As the stripe count is increased, spreading the load over more OSTs, throughput
increases. Single stream writes top out above 400 MB/s and reads exceed 700 MB/s.
Figure A-4 compares write performance in three cases. First is a single process writing to N OSTs,
as shown in the previous figure. Second is N processes each writing to a different OST. And
finally, N processes to different OSTs using direct I/O.
Figure A-4 Single Client, Multi-Stream Write Throughput
For stripe counts of four and above, writing with separate processes has a higher total throughput
than a single process. The single process itself can be a bottleneck. For a single process writing
to a single stripe, throughput is lower with direct I/O, because the direct I/O write can only send
one RPC to the OST at a time, so the I/O pipeline is not kept full.
For stripe counts of 8 and 16, using direct I/O and separate processes yields the highest throughput.
The overhead of managing the client cache lowers throughput, and using direct I/O eliminates
this overhead.
The test shown in
Figure A-5 did not use direct I/O. Nevertheless, it shows the cost of client cache
management on throughput. In this test, two processes on one client node each wrote 10 GB.
Initially, the writes proceeded at over 1 GB/s. The data was sent to the servers, and the cache
A.2 Single Client Performance 65