Attendees: Philippe, Danilo, Zhe, Oksana, Brian.
Oksana: I was able to reproduce the problem on linux. A ‘tree’ function is calling the CRC function (original version and intrinsic version have different version). The behavior is very stable/reproducable on linux but on MacOS it is unstable (end result changes). And indeed valgrind is not clean on valgrind. I am analyzing the 4 lines of code (or so) that are suspicious.
Brian: In the fallback case, we should not be using the heavily optimized version especially if it is suspicious.
Oksana: To use the intrinsic version on MacOS we are missing the ‘architecture’ based trampoline function supported by the compiler.
Brian: Proposal mode until May 1st.
Danilo: Finalizing the feature for the release. I was able to read CMS opendata with RDataFrame.
Guilherme [Danilo reporting]: Not much new on Parallel Buffer.
Brian: Did we give you any of the NanoAOD data files?
Danilo: Not yet, we are very interested (of course :) ). Especially since the regular data file require dictionary for full reading.
Brian: Andrea Ricci might be able to give you some information.
Danilo: DataFrame vs BulkIO and overhead. TDataFrame ‘just’ filling a data frame from ‘just’ created data (i.e. no I/O involved). TDataFrame’s overhead is of the order of magnitude as the histogram filling.
Danilo: I will look into the TBranch::SetAddress probably next week.
Zhe: RAS.
Philippe: Significant enhancement to the TTreeCache FillBuffer algorithm.
Brian: What is the guidance on the TTreeCache, still the same or different?
Philippe: Yes, it is just sturdier and more stable :). If you want to stay within a strict budget, set the cache to half of the cluster size. The rules is as follow: Use up to 4 times the requested amount to fit at least one basket per branch, Use up to 2 times the requested amount to fit the first cluster. Use only up to the requested size to fit extra cluster.
Brian: Friday the 11th, I will be at Fermilab.