[PaulClayton]

Hi,

Our configuration two VIOs(1.5.1).
Each VIO has two 2G 5773 fibre cards. We use SDDPCM MPIO drivers.
The back end disk is an EVA8000 with an IBM SVC inbetween. The whole fabric runs at 2Gb.
Test I have run are as follows.
Disk write test using VSCSI device to SAN on DLPAR - 16MByte/s
Disk write test using SAN attached disk on VIO - 27MByte/s
Disk write test using SAS attached local disk - 55MBytes/s.
Read are double.
This server is not yet active or in production.

I tested our old P5(1.6GHZ) CPU which is very busy and not optimally configured. 30Mbyte/s

I have check that the queue_depth parameter is identical, and set the various paramemters according to the best practices documents. The DLPAR is run 6.1 SP4.

Two issues appear here, that is that I am loosing about 45% across the hypervisor, and there appears to be an issue between the P6-SVC-EVA storage.
Any thoughts on where the problem could be.

[cdelgadop]

Hi

Yes, it's possible. Remember you're going through another layer of abstraction when using VIOS and one more one when running through IBM SAN Volume Controller. It's supposed you have already run the compatibility matrix and you have installed the most updated drivers and microcode levels for your adapters, as well as your SAN team have already checked zoning and any patch level SAN components might need.

I would deinstall SDDPCM and install SDD, i think SDD's algorithms are more efficient than those in SDDPCM.



Hope this helps

[john.samons]

I assume the p5 & p6 are both connecting to a EVA8000. Is the disk configuration on the EVA8000 the same for both? What is the configuration if you please. RAID, parity groups, LUN sizes, etc..

I would also think that your performance should be quite a bit more than what your getting now.

I'd suggest doing performance tests from the VIO server first until you've found reasonable performance.

How are you performing your test?

How is the disk from the EVA8000 presented to the VIOS, and how do you present it to the LPAR?

[PaulClayton]

After I changed a few block sizes, I obtained better performance. Optimum seemed to be 8192, which gave me 70/45 Mbytes/sec over physical/virtual scsi. I got some information from IBM, that pretty much confirmed these performance figures. Initially I was using 1K blocks to test with.

[cdelgadop]

OK

That's good. Remember your VIOS is connected to a SVC so you might need to do some tuning at that level too.

SVC from IBM seems to be a good solution to make all your storage be seen with a single driver but at the same time is another layer data has to pass to get to the physical disk. Those many layers we are using nowdays add some latency to the process and i think we are going to start seeing many performance problems with SVC.

[john.samons]

Great! I'm happy you got better performance. Maybe do a couple more tests, you might be able to get a bit more out...

I agree with cdelgadop, the SVC does affect performance in a large way. I don't easily recommend it for customers, except if it's really needed. But then again, you can design your SAN differently...

I'd recommend... SAN Volume Controller: Best Practices and Performance Guidelines
IBM Redbooks | SAN Volume Controller: Best Practices and Performance Guidelines

[PaulClayton]

Got very fustrated, and went one better. I trashed my vios and client partition, and re-initialised the P6 managed system, and rebuilt the vios and client partition.
Now I am getting 220MBytes/second throughtput, through a 4 disk stripe with a 1Mb stripe width.
I wonder if the fact that the server was initially factory configured, whether this had not played a part in the problem, and the re-initialization resolved the problem.
The results are dramatic, and now a number of other items are now working that did not work well, including network and fabric failover.
Thanx for you inputs.

[ross.mather]

Out of curiosity what was your write test? I know Oracle always advises a stripe size of 1Mb as that then links in with the way Oracle is written. Maybe its related that way?

[john.samons]

Now that's interesting! Congrats with the performance! Told you you'll get better performance out of that EVA...

I wasn't aware your systems were factory installed...

Out of lots of previous little funnies and gremlins with pre-loaded LPAR's (with or without VIO) - I've always adapted the standard to reinstall FRESH with my own media and my own settings. Works every time!

NIM still seems to give problems sometimes, with the bootpd but other than that it is fine....

FROM:
Test I have run are as follows.
Disk write test using VSCSI device to SAN on DLPAR - 16MByte/s
Disk write test using SAN attached disk on VIO - 27MByte/s
Disk write test using SAS attached local disk - 55MBytes/s.

To: Now I am getting 220MBytes/second throughtput

Seems like your performance improvement from the initial tests has increased 92.72% - 75% Nice!

[cdelgadop]

Hi

Well, by default i always reinstall all AIX/Linux LPARs. Place a call on IBM and, i think it's a CMT, and make a formal complaint about this issue. Sometimes we as customers have to complaint to our providers so they can detect problems and correct them.