PROOF CPU Efficiency

ROOT
1

Hi Gerri, all,

We would like to estimate the CPU efficiency of each PROOF job running on the worker nodes. If this value is low, it could give us a hint on weather we are overloading the storage and there is no gain on increasing the number of parallel slots assigned to PROOF. Of course, only if our jobs are more I/O bounded than CPU bounded (which is mostly our case). Is there an easy way this value could be found from PROOF?

Thanks,

Isidro

2

Hi Gerri, all,

we are still trying to measure the CPU efficiency of each PROOF job running on the worker nodes.
We found that there are some time histograms defined at [1]:
PROOF_ProcTimeHist and PROOF_CpuTimeHist, from them should be
possible to extract the efficiencies.

We are able to access the histograms via:
gProof->GetOutputList()->FindObject(“PROOF_ProcTimeHist”);

However, we did not succeed in understanding the information.

Would you know how to extract those times per worker?

Thanks in advance,
Ana.

[1] root.cern.ch/drupal/content/crea … mance-tree

3

Hello, I’m not very expert, but on our cluster we have estimated the performance (only cpu or I/O) with TProofBench, see here: root.cern.ch/drupal/content/proo … proofbench

4

Thanks for your reply.

We are interested in computing the CPU efficiency
of some of our PROOF analysis, so we would like
to know which is the CPU time that our analysis
takes to run as a function of the number of workers.
We do manage to extract the wall time from the
processing time but not the CPU time.

It seems those two benchmarks get the events per second
instead of CPU times. Moreover, we are not sure on how
to adapt them to our PROOF analysis.

Thanks,
Ana.

5

maybe this one: root.cern.ch/drupal/content/proo … veSelector

[quote=“arodrig”]Thanks for your reply.

We are interested in computing the CPU efficiency
of some of our PROOF analysis, so we would like
to know which is the CPU time that our analysis
takes to run as a function of the number of workers.
We do manage to extract the wall time from the
processing time but not the CPU time.

It seems those two benchmarks get the events per second
instead of CPU times. Moreover, we are not sure on how
to adapt them to our PROOF analysis.

Thanks,
Ana.[/quote]