ROOT I/O Meeting

Friday 24 Mar 2017, 15:55 → 17:20 Europe/Zurich

VIDYO

Attendees: Viktor, Brian, Philippe, Zhe

Brian: I made some progress on TTreeReader fast getting close to be able to run performance test; hopefully result next week.

Zhe: TBB parallel unzipping and comparing to original design.  So far it is slower … original 1s per 1000 events.  TBB 9s per 1000 events.  40% of the time is spent in ‘receiveAndSteal’ TBB function.  But this does not account for all the slowdown.

Brian: I suspect the average task size is too small (too much ‘task’ switching).

Viktor: Running SparkRoot on CERN IT-DB on an Intel Cluster. 14*2*18 cores. Data: Muonia CMSSW AOD 2010, 1.2TB.   Spark provides some monitoring (performance measuring) information but some of it is confusing, in particular for the raw I/O.

Viktor: Some oddities in the performance including a sharp decrease in performance when going over one task per core (followed by increase in performance as the number of task increases) and an (somewhat linear) increase in total CPU time when increasing the number of threads (however this increase is not seen when the number of threads per machine is constant and the number of machines increases).

 

16:00 → 16:20 Update on running spark-root with cmssw aod on Intel’s cluster

Speaker: Viktor Khristenko (The University of Iowa (US))

sparkroot_performance_24032017.pdf

16:20 → 17:00 Round Table