CPU monitoring in an lpar environment

[andy61]

Hi,
I'm looking at the best way of checking CPU in terms of the effect on performance on our p570s with uncapped lpars. From what I can see, there are three limits that will have an impact
Overall usage
Shared Processor Pool Usage
An lpar reaching its VP limit.

I've used lpar2rrd for the first two, and from the same source on the HMC, have developed Patrol KMs. What I see is CPU resource issues when these values are above 80-85%, which seems quite a low threshold.
Looking further, I set up vmstats running on each lpar in 1 sec intervals to record processor usage and wrote a script to harvest the data. From this I see a much higher level of usage - though i think this method is more suspect.

What I'd like to know is has anyone else noticed such behaviour ? Is it the case that in the virtual environment our monitoring thresholds should be set lower than they would on dedicated proc lpar ?

 

[w5000]

try nmon and/or lpar2rrd

 

[AIXLogician]

Yes,
Nmon is the very good tool. Not only for CPU but for memory, IO, Network etc.., etc.

Also try using sar and mpstat

SARFARAZ AHMED SYED,
Sr. Systems Engineer

 

[andy61]

Thanks for the postings - I've been wrestling with another problem hence my lack of a reply.

I'm already using lpar2rrd and its results are similar to my Patrol KM - which is likely as they both use the same source.

I use nmon a lot, but I need frame wide stats on CPU ie overall and Shared Processor Pool usage. nmon does have some capacity for SPP stats, but I think it can only show how much of the pool the lpar is using.

I think what I need to establish is what is the threshold for usage as reported by lapr2rrd/patrol when cpu issues are likely. Experience indicates that either this threshold is low or the figures are inaccurate. Has anyone else seen this ?

I'd like to use topas -C, but it requires xmtopas and certain corporate restrictions will make that difficult.