EVA 4400 performance

[Jazzh_1]

I’m doing some performance analysis on a new EVA 4400. I’m the only one who is using the system right now. The benchmark is sysbench. http://sysbench.sourceforge.net/. The server is DL380 G5 (RedHat 5.0) connected via MDS cisco switches to EVA 4400. Disk group is 32x146G, 15K rpm disks. The HBAs are Emulex 4Gb cards. The test I’m running is
# ./sysbench --test=fileio --max-time=60 --max-requests=1000000 --file-num=1 --file-extra-flags=direct --file-fsync-freq=0 --file-total-size=128M --file-test-mode=rndrd run

This test makes single threaded random reads for 1 minute from 128Mb file, created early, with 16K blocks (this is the page size for MySQL). The extra-flags=direct bypasses the operating system cache and makes direct access to the files system (similar to MySQL O_DIRECT configuration). The result I’m getting is around 3,300 IO/Sec. This is lower than I would expect and comparable to the result I got on Sun x4440 server with 8 internal 15K RPM drives. I would expect the numbers to be above 4,100. There is another problem that puzzles me, At some point during the test, I would expect the entire 128Mb file to be loaded into EVA cache and I would expect the performance to be what the OS can get out of EVA with 100% cache hit. Somewhere around 13K IO/sec. Any consecutive runs of the same command should also give me 13K IO/sec since I’m the only one who is using the system but I still get the 3 thousands and something IO/sec.

Am I wrong in my assumptions? Is there a configuration/tuning I should pay attention to?

P.S. This thread has been moved from HP 3PAR StoreServ Storage to Storage Area Networks (SAN) (Enterprise). -HP Forum Moderator

[IBaltay]

Hi,
perhaps the EVA 4400 performance white paper could help you:
http://h71028.www7.hp.com/ERC/downloads/4AA1-8473ENW.pdf

[Trever Furnish]

I'm doing similar work, although my time is quite constrained. I'm wondering, Jazzh, whether you were able to get better throughput than what you reported above, and whether per chance you documented all of your results and could share those?

[Víctor Cespón]

With 32 x 15K drives and 100% reads, you should be able to reach 5200 IOPs before the latency goes above 15 ms.
You would need to capture data with EVAperf to compare the requests that the hosts sends to the EVA with the requests processed by the EVA.
Could be a limitation of the HBA driver, there was a case in this forum recently.

[Jazzh_1]

I did try the same test on 3par and got the same result. The interesting part about this is that running sysbench with these parameters, file-total-size=128M file-num=1, means that read is done from one 128MB file. The file is small enough to be loaded completely into EVA cache (at that time I was the only one who was using the system). I would expect results that are much better than the number of spindles allow and something close to results that are usually achieved if you allow the OS caching (the --file-extra-flags=direct prevents use of OS caching). I was trying to repeat the 1st test from here http://www.mysqlperformanceblog.com/2008/03/05/evaluating-io-subsystem-performance-for-mysql-needs/

[IBaltay]

Hi,
the problem may lay in random reads, which cannot get much use from the cache prefetch/read ahead and most of the time they need to go to the disks itself... and 32 spindles could not be optimum enough to "go closer to the ns instead of ms".....

[IBaltay]

You may try to tune it with:
a) increasing the IO queue length (throttle)
b) disk allignment
c) Linux device mapper ALB settings

[Jazzh_1]

I know about the first two and did play little with the settings. Can you please elaborate on the 3rd one?

[IBaltay]

FEW DOCS:

http://h20000.www2.hp.com/bizsupport/TechSupport/Document.jsp?lang=en&cc=us&objectID=c01475580&jumpid=reg_R1002_USEN

http://h20000.www2.hp.com/bizsupport/TechSupport/Document.jsp?lang=en&cc=us&objectID=c01475612&jumpid=reg_R1002_USEN

DOWNLOADS:
http://h20000.www2.hp.com/bizsupport/TechSupport/SoftwareIndex.jsp?lang=en&cc=us&prodNameId=3559652&prodTypeId=18964&prodSeriesId=3559651&swLang=13&taskId=135&swEnvOID=4006

Release Notes:
http://h20000.www2.hp.com/bizsupport/TechSupport/SoftwareDescription.jsp?lang=en&cc=us&prodTypeId=18964&prodSeriesId=3559651&swItem=co-66308-1&prodNameId=3559652&swEnvOID=4006&swLang=13&taskId=135&mode=4&idx=0

Migration to the device mapper:
http://h20000.www2.hp.com/bc/docs/support/SupportManual/c01388403/c01388403.pdf?jumpid=reg_R1002_USEN

Device mapper resource page:
http://sources.redhat.com/dm/

[IBaltay]

one more doc/download :-)
http://h20000.www2.hp.com/bizsupport/TechSupport/SoftwareDescription.jsp?swItem=co-64600-1〈=en&cc=us&idx=0&mode=4&