One of the most popular post on my blog, is the one about monitoring a CPU by CPU cycles dispatched by a POWER Hypervisor link.
With POWER8 we receive new attributes in lslparutil command.
One of the biggest disadvantage of the lslparutil command so far, was a limitation if you you an Lpar in a dedicated processor mode. When you run an lpar in dedicated mode (capped), you basically want to ensure that the lpar always get as much CPU cycles as it needs, without any delay in a CPU dispatch. The problem is that you couldn’t see how much CPU partition really uses from the Phyp perspective (of course you could see this from the OS level). For instance, if you use LPAR2RRD tool, you didn’t know if the Lpar already reach the limit or not (you can’t do any planning).
With POWER8 IBM comes with new attributes for lslparutil command, which in my opinion allows you to track the CPU utilization even when the Lpar runs in dedicated mode.
These attributes are following:
- total_instructions: The number of instructions performed by the partition since the managed system was started. It is independent of the whether the partition is in its idle loop or running real work, the instruction count will be incremented as instructions are completed.
- total_instructions_execution_time: The amount of time instruction counts were collected since the managed system was started. The time value also is not gated by the run latch and is a measure of the time the partition was running on a physical processor.
So the lslparutil command gives you now a following output:
lslparutil -r lpar -m POWER8 –filters lpar_ids=1 -n 1
time,event_type=sample,resource_type=lpar,sys_time,time_cycles,lpar_name,lpar_id,curr_proc_mode, curr_procs,curr_sharing_mode,curr_5250_cpw_percent,mem_mode,curr_mem,entitled_cycles, capped_cycles,uncapped_cycles,shared_cycles_while_active,idle_cycles,total_instructions,total_instructions_execution_time
I think you can do the math:
AvgtimePerInstruction = total_instructions_execution_time / total_instructions
There is a lot of questions in this post, because I don’t have an access to a POWER8 anymore. Therefore, I can’t give you precise information. As soon as I get, I’ll definitely publish them.
2 thoughts on “Monitor CPU, POOL utilization from PHYP cycles on POWER8”
Hi Bart, great article. To simplify manual labor to collect data and perform such kind of performance monitoring I would suggest seriously to get familiar and use PowerVP, where all the necessary views of the logical and physical structures on a POWER server are well collected and also displayed. http://www.youtube.com/watch?v=xL9E7-gVw4Q
Thank you. Yes the PowerVP simplify many monitoring aspects, but it requires very recent firmware levels on the server and the HMC. It will take some time, while people will run their machines on the proper level.
I hope to cover the PowerVP topic soon on this blog.