HP StorageWorks Scalable File Share - Manual (Page 53)

Another way to measure throughput is to only average over the time while all the clients are
active. This is represented by the taller, narrower box in
Figure A-8 . Throughput calculated this
way shows the system's capability, and the stragglers are ignored.
This alternate calculation method is sometimes called "stonewalling". It is accomplished in a
number of ways. The test run is stopped as soon as the fastest client finishes. (IOzone does this
by default.) Or, each process is run for a fixed amount of time rather than a fixed volume of data.
(IOR has an option to do this.) If detailed performance data is captured for each client with good
time resolution, the stonewalling can be done numerically by only calculating the average up to
the time the first client finishes.
NOTE: The results shown in this report do not rely on stonewalling. We did the numerical
calculation on a sample of test runs and found that stonewalling increased the numbers by
roughly 10% in many cases.
Neither calculation is better than the other. They each show different things about the system.
However, it is important when comparing results from different studies to know whether
stonewalling was used, and how much it affects the results. IOzone uses stonewalling by default,
but has an option to turn it off. IOR does not use stonewalling by default, but has an option to
turn it on.
A.6 Random Reads
HP SFS with Lustre is optimized for large sequential transfers, with aggressive read-ahead and
write-behind buffering in the clients. Nevertheless, certain applications rely on small random
reads, so understanding the performance with small random I/O is important.
Figure A-9 compares random read performance of SFS G3.0-0 using 15Krpm SAS drives and
7.2Krpm SATA drives. Each client node ran from 1 to 32 processes (from 16 to 512 concurrent
processes in all). All the processes performed page-aligned 4KB random reads from a single 1
TB file striped over all 16 OSTs.
Figure A-9 Random Read Rate
For 16 concurrent reads, one per client node, the read rate per second with 15K SAS drives is
roughly twice that with SATA drives. This difference reflects the difference in mechanical access
time for the two types of disks. For higher levels of concurrency, the difference is even greater.
SAS drives are able to accept a number of overlapped requests and perform an optimized elevator
sort on the queue of requests.
A.6 Random Reads 53