IOPS using esxtop, readout appears inaccurate

[udp666]

Hi everyone,

I have a dedicated VM on my esxi server (4.1) that I need to provide IOPS information to my data center team.

They will use IOPS information to determine SAN requirements.  I did some digging around and found the esxtop command and the CMD/s output.

I have conducted a series of tests and wrote the esxtop output to a file.  After examining the file in more detail, it appears the information is not accurate or I may be missing something.

My tests are as follows:

- Send 100 Events Per Second

- Send 500 Events Per Second

- Send 5,000 Events Per Second

- Send 15,000 Events Per Second

These events are both binary and ascii format that are written to two databases across a RAID 5 array of SATA 7200 RPM disks.

After looking at the output of CMD/s, the 5K EPS has higher numbers than what is reported with 15K EPS.  This cannot be correct.  Also, sometimes the number is much higher than what I ever think the disk is capable of reporting. IE: 100+

I set the interval to 2 seconds.

My question is, am I doing something incorrect and is there a more accurate way to capture the IOPS of the VM?

Thanks,

Dan

[MKguy]

The CMD/s metric is not necessarily the best way to measure IOPS from a system:

http://www.yellow-bricks.com/2010/06/24/cmdss-vs-iops/

Overall, the esxtop statistics do reflect accurate, real numbers. If you check IO statistics on the storage array side, the numbers should match up very closely.

I'm not sure what your test is doing exactly. When you refer to "events per second" you're not expecting to translate this to IOPS directly, are you?

Another really cool way to analyze IO metrics from certain VMs is the tool vscsistats:

http://communities.vmware.com/docs/DOC-10095

http://communities.vmware.com/docs/DOC-10084

http://communities.vmware.com/docs/DOC-10104