EVA4400 performance documentation

[Xavi.e.r]

Could anybody refer me to an HP documentation where I can find information on general IOPS calculations/sizing and respective calculations depending on vRAID 1,5,6?.

I already checked these ones but did not find it
http://h20195.www2.hp.com/v2/GetPDF.aspx/4AA2-0994ENW.pdf

EVA4400 performance white paper 4AA1-8473ENW

I have seen different ITRC posts suggesting a general estimation of 170-200 IOPS for 15K drives and 120-150 for 10K drives.

I am trying to determine a performance baseline of an EVA and to know whether is overloaded or reaching max performance capacity with the current configuration.

EVA4400 w/ dual controllers (09534000 firmware)
4GB total cache
2x Embedded 8GB switches
24x 450GB 15k & 10x 146GB 15K drives.

[Peter Mattei]

Xavier

Have you seen this support document?
http://h20000.www2.hp.com/bizsupport/TechSupport/Document.jsp?objectID=c01685240

I think together with the "A tactical approach to performance problem diagnosis" paper you referenced above you do have the info you need to get it sorted out.

Cheers
Pete

[Mikko Niskanen_1]

See e.g. this thread:

http://forums13.itrc.hp.com/service/forums/questionanswer.do?admit=109447627+1289562717213+28353475&threadId=1262271

[Xavi.e.r]

Thanks for the input.
Now, should I assume that the 120 and 170 IOPS per disk applies the same way when having VRAID 1,5,6 or how do I figure it out?

Also, about the read and write latencies, should I consider them individually each host port on the EVA or should I do a sum of the reads and a sum of the writes (EVA4400 has 4 ports: 2 ports per controller)

I did a capture with evaperf for a few days, using tlviz, I see in the array an average of 3000 IOPS (total req/s) and peaks on the 6000 IOPS. I see each host port very similar, the read latency averages 20-25ms and write latency 8-16ms. The latencies are over the numbers that HP recommends but the IOPS within the limits from the 5780 max calculated IOPS (2 disk groups, 1 with 24 450GB 15K and the 2nd 10 146GB 15K (1 vraid 5 and the rest vraid6) . Is this still an indication that EVA is overloaded or what other counters should I take in consideration? I am kinda confused here. Thanks

[Peter Mattei]

The 120 and 170 IOPS are the backend performance numbers per disk you will find in EVAPerf.
Depending on the vRaid level these will be reached with more or less front-end IO.

Pure reads are equal for all vRaid levels.

Example for 1000 IOPS random front-end writes:
on vRaid1 this will caues 2000 back-end IOs
on vRaid5 this will cause >2000 back-end IOs
- Read the original data and parity block (two requests)
- Calculate the new parity block
- Write the new data and parity block (two requests)

For vRaid 6 it will be even more.

Cheers
Pete

[Mikko Niskanen_1]

Aye,

If it is IOPS you're after, i.e. for an OLTP-style database use:

- As many 15k disks as you can afford
- All disks to one disk group
- Total # of disks divisible by eight i.e. 8, 16, 24, 32,...
- Use only vRAID-1
- check SCSI command queue length


Suggested reading EVA Best Practises docu:

http://h20195.www2.hp.com/V2/GetPDF.aspx/4AA2-0914ENW.pdf

[Xavi.e.r]

Thanks for the responses. Now based on the previous info I included in the post above, Should the latencies really be under the those numbers (less than 5ms for write and 15ms for reads)? Would these latencies mean that there's something that needs to be done to reduce them ? if so what would it be? more drives?

[Mikko Niskanen_1]

For Oracle, latencies of x < 10ms are okay, 10 ms < x < 20ms are warning signs and latencies x > 20 ms are not good, you're about to hit performance ceiling.

If you check Fig 1 in http://h20195.www2.hp.com/v2/GetPDF.aspx/4AA1-8473ENW.pdf, the read performance slopes nicely but you should be concerned if you have heavy writes, then you could hit a performance wall. Depends on your read/write ratio, if it's mostly reads, then 4400 will do good, but if you experience high write latencies over a long period (i.e. apart from occasional spikes), then you're in trouble.

[McMac]

Hey Xavier,

I've spent the last few weeks doing just what your looking at and have written a tool. If you can get a perfmon log to me I reckon I could spot any performance bottle necks in about 10 minutes.

I'm interested in testing the tool on other peoples machine's perflog data, so we'd both get something out of it.

[Xavi.e.r]

Hey McMac
Thanks for the offering, I have a evaPerf capture that it's about raw 650MB/zipped 15MB.

Let me know what do I need to do and what will your tool will do as well.

Thanks