MSA1000 read performance problem

[ianvt]

We have a customer who is experiencing SQL read performance problems. A SHOW PERF LOGICAL shows that the average write latency is between 20 and 22 and that the average read latency is between 75 and 78.
What can be done to resolve this? Will it help to upgrade the cache from 256MB to 512MB?

[marsh_1]

hi,

it might depending on how your app is working read and write wise, can you post up more details on the setup of the raidsets etc any other systems using the msa that'll help with any suggestions people might have.

[ianvt]

They have one ADG raidset shared by 2 VMWare servers. A total of 20 VM servers, both Windows and Linux, are running on the 2 VMWare servers, with 3 running SQL.

Hope this helps.

[Víctor Cespón · Accepted solution]

How many disks there are on that MSA?

You cannot run 30 servers (the fact that they are virtual does not change the amount of I/O they need) and a SQL server from a box with 4 disks.

We never had heard of performance complains on MSA until customers started virtualizing and where previously they had 5 servers now the want to run 30 or 40 virtual machines on the same hardware. That's impossible, rotating-platter disks have a limit to the amount of I/Os they can sustain.

[ianvt]

There are currently 14 disks. We are going to add another MSA30 with another 14 disks.

[marsh_1]

hi,

it might not help you that much in the long term to add more disks , the msa1000 is pretty good for an entry level array but there is only so much you can put through it and it is now retired. if you are going to continue to add to your virtual hosts i would suggest you look to upgrade your storage.

[Richard Tengdin]

After you add the new drives I'd also recommend converting the RAID6 storage to RAID10. As it is you have to read the entire 14 drive stripe then write to data data, Parity1 and Parity2 disks to handle one write I/O from one server. Making everthing RAID10 means you write the same data to two disks, and the rest of the disks are free to service their own local I/Os.