EVA8100 Performance Pro

[Mr. Magic]

I have a question about performance. I am seeing very bad perofmance on one of our EVA8100 which has 108 FC Disks and 36 FATA Disks. It is used for VMware and SQL storage. I am seeing many high latencies and phsical queue depth goes many times much higher then 2. I am capturing PDG counters from evaperf. I have 2 of those EVAs and one with the problem has less servers on it (assuming smaller load). The only differene I can see is this:
EVA with a problem physical disk layout:
14 Enclosures same configuration:
1 2 3 4 5 6 7 8 9 10 11 12 13 14
FATA FATA FCFCFCFCFCFCFCFCFCFCFC FATA

Would physical location of disk impact perfoamnce of disk Group? All FATA disks are in one Virtual Group.
SAN was checked for configuration issues and nothing was found. Any ideas?

[Víctor Cespón]

No, physical layout of the disks does not affect performance significatively, not on a EVA8100, that has CTS I/O modules.

The FATA disks are 1 TB? Can you check their temperature?

Also, you need to check how many I/Os the FATA disk group is receiving, here's a document stating how many I/Os per disk you can expect:

http://h20000.www2.hp.com/bizsupport/TechSupport/Document.jsp?objectID=c01671044

[shIVinator.]

Your setup looks like it is as optimal as ou can get. Use *vcespon* link above as a general guideline for IO because the amount of IO that a physical disk can do is also dependent on the the type of IO. e.g. small random vs. large sequential.

As for the performance of the VMware environment you have to remember that VMware does a terrible job load balancing. First thing that I would look at would be the host port counter to see if the load is balanced across all ports:
evaperf hps -cont
That could be a bottleneck right there if you are pushing too much IOPS or MBs over one single link.

If that is the bottle neck then look at how many hosts there are on the datastores. There should be a max of 15 VMs per datastore.

One last note would be to not use the FATA drives for any type of high IO. FATA is meant as low cost low utilization solution (archive data...).

Duane

[Mr. Magic]

Thank you for all replys. As far as temperature goes does it apply to FATA only? I know that HP had issue with firmware which caused some performance problems if drive went above certain temperature but we have latest firmware which fixed this problem. We are creating smaller LUNs for VMware as we had big 2TB LUNs and I can see that performance is getting better but still not the best. My next question is to multipath load balancing on vmware, as mentioned above vmware does terrible job but I see Write Latencies and Read Hit Miss Latencies well above acceptable treshholds (evaperf PDG and VD counters). In my view if path would be overloaded disk groups would not be affected?

[shIVinator.]

Good thing you got off that big LUN, that's a killer on VMware perf. If the path can't handle the IO being push through the pipe then there will be a problem on the PDG as well. I have seen this before and the first place I always check is from the evaperf command line and use:

evaperf.exe hps -cont

you will immediately see if there is a problem.

Duane

[Mr. Magic]

Now i did capture counters with hps and they look quie ugly on a controler. Write latencies on one port are up to 15k ms :( If spread load on a FPs would that improve my performance? I added listtle extract from my evaperf counter.