Read / write latency - EVA6400

[inex]

What is acceptable read and write latency on an EVA6400?

-inex

[Amar_Joshi · Accepted solution]

Inex,
It also matters where you are seeing the latencies?

If seeing from the EVAPerf, you will get different values for physical disk and diskgroup. For the sake of discussion I will keep the physical disks out of perfromance discussions because write cache on HSV controllers and read-ahead (cache hit) will play a major role in performance enhancement.

From the DiskGroup view, 15-20ms should be reasonable but depending upon applications (like fileshares, backups) 30ms will be acceptable.

A major factor in reading performance logs is collection interval, if collecting at very high interval (e.g. 60 seconds) a lot of peaks will be averaged out and will make performance study difficult.

Hope that helps.

[inex]

Thanks. How do you differentiate on read and write i/o's?. For ex. is 20ms ok for read and 5ms ok for write?

-inex

[Víctor Cespón]

Check this article:

http://h20000.www2.hp.com/bizsupport/TechSupport/Document.jsp?objectID=c01685240

The commonly used values are:

Read latency
< 15 ms for FC disks
< 25 ms for FATA disks

Write latency
< 5 ms

Also read these:

http://h20000.www2.hp.com/bizsupport/TechSupport/Document.jsp?objectID=c01846577

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

[inex]

I'm running CA and mirror data from the EVA6400 to a EVA4400. There are 11 luns of size 1500-2000GB. When I started syncronizing the last 2 luns, the write latency on the lun owing host ports went up to about 150ms. I had one lun on each controller. Both started with sync mirroring, then I moved one of them over to async and the write latency went down on the controller ports and that lun managed to sync a lot quicker than the other lun which was doing synchronous. And the write latency was again very high on those controller ports owing that lun.

-inex

[Amar_Joshi]

Inex,
In the mentioned sync/async scenario I couldn't figure out where were you measuring the latencies? Was it on LUN being synchronized or all the LUNs?

High latencies on synchronous replication is normal because it will include the delay to transmit the data to remote array and acknowledgement from the remote array. If the link between two arrays is not very fast it's highly recommended to use asynchronous replication.

[inex]

I measured the latency on the EVA6400 front end ports while I started to sync 2 new luns to the DR groups, which already held 9 luns which were running in sync mode. The latency was very high when trying to include these two with sync mirroring.

-inex