Hi;
After MD3000i performance issue, I need to calculate the IOPS usage over 5 ESX 3.5 connected to the ISCSI bay.
I don't want to benchmark it, but just to agregate the load and have the total IOPS usage .
Is there a good counter into ESX (or VirtualCenter) that can give me an overview (value or trends) of it ?
I read those links, but no direct info related to IOPS value.
Esxtop manual
http://www.vmware.com/pdf/vi3_301_201_resource_mgmt.pdf#page=159
Esxtop Performance Counters
http://communities.vmware.com/docs/DOC-5240
Storage Performance Analysis and Monitoring
http://communities.vmware.com/docs/DOC-5490
Can you help ?
Thanks
Js
IO counts are reported in VC as disk.numberRead.summation and disk.numberWrite.summation. These can be converted to IOPS by dividing by the sample window (which is 20s in real-time stats.) See for more info. I've also updated the esxtop wiki document you referenced to include the equivalent esxtop counters.
Scott
Can someone please verify the correct method for calculating IOPS with VirtualCenter 2.5. If I understood this post correctly, then IOPS can be calculated by dividing "Disk Read Requests" and "Disk Write Requests" by the data sample rate.
In my attachment the data sample rate was 120 seconds. So does this mean that the average IOPS in the attached example is 2.97 (3.65 + 352.58 / 120 seconds)?
Your math is good, but I think that your sample rate is bad. Unless you know better, the sample rate should be 20s for real time data.
Scott
The report sample (IOPS.PNG) was not from "Realtime" data, it was older data. Therefore, if I understand this correctly, the sample rate should equate to the relevant Statistics value set in "Advanced\VirtualCenter Management Server Configuration\Statistics", which in our case is 2 minutes (or 120 seconds). Is that correct?
See StatCollectionSettings.PNG attachment for screen shot of our Statistics settings.
Is the same true of ESXTOP stats? If I run esxtop in batch mode collecting statistics every 10 seconds, should I divide the Commands/sec value by 10 to get the true IOPS value?
If you found this or other information useful, please consider awarding points for "Correct" or "Helpful".
Please visit for News, Views and Virtualization How-To's
No math is needed for units whose results are reported "per second". The rate you are seeing has already been normalized for a one second window.
Scott
More information on my communities blog and on Twitter:
http://communities.vmware.com/blogs/drummonds
Thanks Scott! I assumed that was the case from the numbers being reported but wanted to double-check.
If you found this or other information useful, please consider awarding points for "Correct" or "Helpful".
Please visit for News, Views and Virtualization How-To's
IO counts are reported in VC as disk.numberRead.summation and disk.numberWrite.summation. These can be converted to IOPS by dividing by the sample window (which is 20s in real-time stats.) See for more info. I've also updated the esxtop wiki document you referenced to include the equivalent esxtop counters.
Scott
So there is a lot of questions on how to convert vmware stats to IOPS and the best answer seems to be running esxtop in batch mode. That however is a pain as even if you only collect the data that you need those csv files grow quickly.
Anyway so I used the disk.numberRead.summation and disk.numberWrite.summation counters using LucD's script from http://communities.vmware.com/message/1501990 to capture 7 days worth of data. The 1 week stats interval collection is at 30 minutes. Would that just mean that to get the IO for one week I would need to divide the disk.numberRead.summation and disk.numberWrite.summation that I get by 1800?