GoogleBenchmark setup

[vincecr0ft]

Two files have been added. runBinnedTest.cxx and buildBinnedTest3.cxx These are simple ROOT macros that have been provided to the ROOT Benchmarking team. Results on earlier tests can be seen here.

As can be seen the time goes up for multiple cores. this is apparently still true with more cores capping the job time at taking more than 12 hours.

A first solution is proposed to kill lines 57 and 61 (mentioning ActivateStatError()) and run on up to 8 cores.

Once the google benchmark code runs well on the knl with proper scaling this test can be extended to other systems (e.g. on NLESC machines) but in the mean while the code to run the tests can be found here.

[egpbos]

/user/vcroft/public_html/forPatrick/outputAugust09

[egpbos]

Yesterday, we ran a simple test with a modified buildBinnedTest3.cxx that produces just 1 channel and no nuisance parameters. We benchmarked it on up to 8 cores, with and without CPU affinity activated. Results:

Blue lines are measured wallclock time, purple lines are theoretical "ideal scaling" times, i.e. the time it would take if the runtime scaled ideally with number of cores. That it does not is a known problem for this type of likelihood which is being worked on.

[egpbos]

@vincecr0ft what are the other parameters in this run actually, N_samples, N_events and N_bins?

[egpbos]

Judging by the other set of workspaces, I guess it's 1 sample, 1000 events, 10 bins, is that correct?

[egpbos]

For completeness' sake, this plot compares the total and ideal wall clock times (blue and red) to the CPU times. Green is the CPU time of the master process, purple the sum of all the MPFE processes' CPU times and yellow the sum of green and purple (i.e. the total CPU time of the entire minimization process over multiple CPUs).