Choose timezone
Your profile timezone:
Powered by Indico
v3.3.3-pre
Jim: Presentation of the status (doing well but having some mysteries in performance profile) the NumPy bridge.
Brian: Several of the additions seems to belong to PyROOT. What would be your top 3 API addition to PyROOT.
Jim: Passing the function pointer to Numba should be much easier. I am thinking of adding it to PyROOT’s TMethodProxy, I also know how to ‘extend’ Numba to deal with ‘unknown/new’ method more automatically. I am been reviewing the various ideas for a while. Still need more time to put it in a form good enough for a PR.
Zhe: I have some update about parallel unzipping (See slides).
She: TBB GetEntry seems to be wasting a lot of time in the TBB layer (‘ReceiveTask’) likely because the number and/size of the task is too small ….
Liz: Is the file you are using a ‘recent’ CMS files.
Brian: It is a bit old and it may or may not explain why the parallel GetEntry does not have enough task.
Zhe: I am limiting the number of branches (200 fish)
Brian: You should look at the 5000 branches ranges. But then again 280 is a reasonably representative number of branches for actual analysis.
Brian: What happened to the work-stealing version (to replace the busy waiting)?
Zhe: Actually the problem is the tail end when there is no more work to be stolen and we are waiting for the completion of the last few.
Philippe: We will need to think about the interface.
Brian: I think it replaces the TBB GetEntry.
Philippe: Humm …. The parallel unzipping sounds more performant on this one example but we need to understand both why it is more performant and if it is performance on ‘all’ uses cases and files (or at least a majority of them)
Brian: 2 small questions. What about LZ4? Google Benchmark?
Philippe: I will work with Oksana on this.
Danilo: Waiting on Vassil to come back from vacation.
Brian: I am really interested in using it for the new I/O performance test.
Oksana: Still working on the LZ4 testing update.
Oksana: Also thinking about ways to test the parallel buffer merger with different configuration (on KNL too).
Peter: David Clark’s PR is waiting for Philippe’s input.