Re: [pcp] Difference between: kernel.percpu.cpu.user and kernel.pernode.cpu.user

[Shirshendu Chakrabarti <shirshendu@xxxxxxx>]

Hi Mark,

Please find the statistics gathered below:

[root@pcp-test-shir3 ~]# numactl --hardware

available: 1 nodes (0)

node 0 cpus: 0

node 0 size: 622 MB

node 0 free: 418 MB

node distances:

node  0Â

 0: Â10Â

[root@pcp-test-shir3 ~]# numastat -v

Per-node numastat info (in MBs):

             Node 0      Total

        Â--------------- ---------------

Numa_Hit        Â46805.92    Â46805.92

Numa_Miss          0.00      Â0.00

Numa_Foreign        Â0.00      Â0.00

Interleave_Hit       31.29      31.29

Local_Node       Â46805.92    Â46805.92

Other_Node         Â0.00      Â0.00

Thanks,

Shirshendu Chakrabarti

On Thu, Nov 27, 2014 at 2:04 PM, Shirshendu Chakrabarti <shirshendu@xxxxxxx> wrote:

Hi Mark,

Thanks Âfor the explanation.

I am running on Amazon Linux - 3.4.71-63.amzn1.x86_64

I am running, PCP-3-10.0-1.

Please see its a t1.micro instance.

For more details: http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/concepts_micro_instances.html

I have a few other questions:

1. Does PCP have any different requirements for running on the cloud.Â

2. Do the metrics need to be interpreted differently when running on cloud.

Thanks,

Shirshendu Chakrabarti

On Thu, Nov 27, 2014 at 1:32 PM, Mark Goodwin <mgoodwin@xxxxxxxxxx> wrote:

On 11/27/2014 05:17 PM, Shirshendu Chakrabarti wrote:

Hi PCP team,

Could anyone please respond the question above :)

The difference is the instance domains and counter aggregation.

kernel.pernode is aggregated over all CPUs on each numa node.
On my laptop with 4 CPU cores on one (fake) numa node, we should
expect kernel.all.cpu.user to equal kernel.pernode.cpu.user, and
the sum of the per-cpu counters to equal the same, e.g. :

# pminfo -f kernel.{all,pernode,percpu}.cpu.user

kernel.all.cpu.user
  value 254548080

kernel.pernode.cpu.user
  inst [0 or "node0"] value 254548060

kernel.percpu.cpu.user
  inst [0 or "cpu0"] value 66377110
  inst [1 or "cpu1"] value 61187360
  inst [2 or "cpu2"] value 64051110
  inst [3 or "cpu3"] value 62932480

The mapping between CPUs and nodes is in hinv.map.cpu_node,
where each CPU (instance) is mapped to a node number, e.g.

# pminfo -f hinv.map.cpu_node

hinv.map.cpu_node
  inst [0 or "cpu0"] value 0
  inst [1 or "cpu1"] value 0
  inst [2 or "cpu2"] value 0
  inst [3 or "cpu3"] value 0

In your case, kernel.pernode.cpu.user is zero, which isn't correct.
What platform and PCP version are you running?

Regards
-- Mark

Thanks,

Shirshendu Chakrabarti

On Wed, Nov 26, 2014 at 3:47 PM, Shirshendu Chakrabarti <shirshendu@xxxxxxx
<mailto:shirshendu@xxxxxxx>> wrote:

  Hi PCP team,

  Could anyone in the team point me to any literature which explains the
  difference between a per node metric and percpu metric. The one-liner and
  extended description are absent in pernode metric case and terse in percpu case.

  For example:

  [root@pcp-test-shir3 pmlogger]# pminfo -f kernel.percpu.cpu.user

  kernel.percpu.cpu.user
    Âinst [0 or "cpu0"] value 993240
  [root@pcp-test-shir3 pmlogger]# pminfo -f kernel.pernode.cpu.user

  kernel.pernode.cpu.user
    Âinst [0 or "node0"] value 0
  I was under the assumption that, kernel.pernode.* and kernel.all.* metrics
  are the same. I am clearly missing something important.

  Thanks,

  Shirshendu Chakrabarti




_______________________________________________
pcp mailing list
pcp@xxxxxxxxxxx
http://oss.sgi.com/mailman/listinfo/pcp