Measuring IOPS?

[hutchingsp]

I'm trying to use the HP/ESX sizer tool and would like to be clear on what it is expecting for average/peak IOPs.

When I was sizing a physical exchange box it was a case of assuming a baseline IOPs for each Exchange User and comparing this total requirement against the likely IOPS of your chosen solution.

With ESX is it simply a case of using Performance Monitor to measure Disk Transfers/Sec?

If so would it be logical or physical that would yield the best figures?

TIA.

[spex]

You can not measure your physical esx setup - the console itself is somehow virtualized.

You should use at monitor that runs under that os that you are willing to be used.

I would prefer iometer - since you can specify different io load's

Regards Spex

[hutchingsp]

Sorry poor explanation on my part.

I want to measure IOPS on existing physical Windows boxes to use the HP/ESX sizer tool and I'm unclear which counters I would use in Windows Performance Monitor (some say logical some say physical).

[VirtualNoitall · Accepted solution]

Hello,

Disk Transfers/sec. It depends a bit on your config but phsyical should give you the whole picture. If you have multiple physical "drives" then you will want to add the numbers together.

[hutchingsp]

Thanks for that.

So at a basic level I'll monitor physical transfers/sec as well a total network bytes sent/rec and processor utilization.

So far the sizer is agreeing with my suspicion that "one big box" is enough in terms of performance (ignoring redundancy issues).

[hutchingsp]

Does anyone have any pointers to what sort of assumptions to make about what sort of IOPs different drive types are capable of?

I'm looking at a Fujitsu Fibrecat SX80, which claims up to 100,000 IOPs, which I would assume is based around it having 1gb of cache per controller?

Basically I'm running Performance Monitor on my physical boxes to try and get a feel for average/peak disk transfers/sec but similarly I need to know what the spindles in the SAN I'm looking at are likely to be capable of.

The Fibrecat's take SAS over FC drives, 10k or 15k.

[VirtualNoitall]

Hey,

As a general rule, for random I/O, 10K drives are about 120 I/O per second and 15K is 180. For sequential 10K is about 290 IO/sec and 15K is 430 IO/sec. These are the top end numbers. Expect to see about 80% of those numbers.

RAID types will impact what the total number is for your array. As an example RAID 5 is very good for reading because you can utilize each drive head on each drive of the array. A write though on a RAID 5 array is very expensive as 4 operations are required to write the data and then calculate and write parity.

Hope that helps!

[hutchingsp]

Thanks VirtualNoitall.

My current thinking is that I would want at least one physical RAID1/10 and at least one physical RAID5 array on the SAN so I could split out the IO intensive/less intensive VM's.

Of course it gets interesting when you're considering virtualizing both SQL and Progress