55th ROOT Parallelism, Performance and Programming Model Meeting

Thursday 28 Feb 2019, 16:00 → 17:30 Europe/Zurich

4/S-030 (CERN)

Danilo Piparo (CERN)

Presents: Guilherme, Elisabetta Manca, Enric, Stefan, Sidong, Enrico, Javi, Stephan, Danilo, Philippe, Andrea Rizzi, Pere

Action: Recompile with ReleaseDebInfo flags and check the difference between small and big pages, as well as re run an example with no jitting. If this is always there, we need to activate hugepages in the application.

Action: slide 12 understand where libCling is locking during the event loop.

Realistic Full Analysis

- Philippe: we could make EnableImplicitMT have a developer switch to have 1 worker and make EnableImplicitMT(1) sequential to give the users more perf.

- Page size effect: we see a difference 8% of runtime (!). Now, the cause could be lookup in the code. The culprit could be the jitted code or the code of dataframe analysis. An effect could be the thrashing of multiple threads.We would need the L1 Load.

16:00 → 16:30 Realistic Full Analysis Computation Graphs: RDF Benchmarks and Performance Improvements

Speaker: Guilherme Amadio (CERN)

Flamegraphs

RDF Benchmarks