More accurate VM right-sizing

[stingray75]

Another request...

Would it be possible to increase performance collection for individual VMs?

I'm trying to right-size VMs and need accurate performance metrics from the last 5 weeks. However, I'm finding that because Veeam rolls up the statistics too much, I'm losing the vital small periods when CPU or memory peaks. For example, a VM might work hard at end of month during a batch run, so I need to ensure it has sufficient CPUs and memory during this short period, or it may regularly spike to 80% a few times per hour. Because these stats are being averaged over time, I'm losing these peaks, so 85% CPU usage might become only average 40% after an hour roll up. This means the oversized reports are not accurate enough for us.

If we could target certain VMs and ask Veeam to collect full 5 min perf data for a set period it would really help. I understand this may not be possible due to the way Veeam stores perf data in its database, but I thought it would be worth asking. I know you can increase the setting globally but this would totally kill our database as we have nearly 2000 VMs.

Thanks

Matt

[Vitaliy S.]

Hi Matthew,

Yes, you are correct that it is possible to increase these settings for all VMs which will definitely kill the database. As the your particular scenario what if we address this "issue" from a different angle? I was thinking about introducing new parameter to all capacity planning and right-sizing reports that would account business hours only. For example, if you skip performance data during weekends, holidayы and night time that should also give you more accurate results. What do you think on that? Will it resolve your current problem?

Thanks!

[stingray75]

In some cases that may work but not all, as we have many VMs which run batch processes out of business hours, so they may not do that much during the day but need a lot of CPU/memory during the night or at weekends. Also, at end of month we generally see increases in demand.

[Vitaliy S.]

Hmm...ok, makes sense. Can you please tell me what rollup intervals would you like to use? Be aware that by default we store 5 minute rollups for the past week, and everything beyond that is rolled up into 2 hour data samples.

[fredbloggs]

I know this has been requested before, perhaps peak should always be peak and not averaged out. When rolling into 2 hour rollups it always counts the peak CPU/RAM in that two hours and not the value at that 2 hour point in time. Would be good if it could have 90% percentile as well to remove unusual spikes that would report 100% but it's not really the peak value.

Means the database size would be retained but more relevant information would be available when sizing in this way

[Vitaliy S.]

When rolling into 2 hour rollups it always counts the peak CPU/RAM in that two hours and not the value at that 2 hour point in time.

It's an interesting idea, but what if VM just spiked for 99% of CPU/RAM usage for 1 minute and the rest of the time interval (2 hours) was idling? I've seen that a lot of times in my lab, still thinking how to better address these kind of situations.

[fredbloggs]

realise the spiking is an issue, yes that would break it, that's why i was thinking maybe the 90th percentile for that period or something. this should get rid of spikes but hopefully give a good idea of the peak utilisation of the VM.
Does that sound possible?

[Vitaliy S.]

Let me check that with the dev team, however I was thinking not about 90th percentile for that period but a pure peak value that existed for more than 15 minutes. With this approach random spikes would not be counted.