Attendees: Danilo, Guilherme, Philippe, David A.  Jim P., Zhe.

Zhe: Restarted work on parallel unzipping.

David:  I had recorded 200,000 buffers from ROOT to test hardware and software decompression.   As of note that we are doing a lot of small decompression (a few bytes), so I wanted to understand the behavior in this case.   So I selected 50 case with a range of sizes.   It is clear that small buffer is bad: slower (10kb is a threshold) and in the bytes range (were size actually increase)
I also tested various version of zlib … the one CMS used (cloudfare + support for multiple architecture) performs remarkably well.

Philippe:  Can you make a presentation at the ROOT I/O workshop.

Philippe: To reduce the number of buffers to test, I recommend that you select both a range of size and compression ratio.

Jim: Nothing to report this week.

Guilherme: I am currently benchmarking the in-process parallel merger.  Will do on KNL, laptop, and server 18Mb.
with pythia and random number generation.  Measure scaling, etc.   So far good performance, the merger not being the bottleneck.
256MB per buffer down to 16MB time goes from 36s to 26s. (drand, 1GB)

Danilo: It looks like we need a SnapShot configuration to allow the user to customize the output file.

Danilo: SaveSelf did not work-out.  How so?

Philippe: Yes, and we ought to switch from testing on size to testing on multiple of GetAutoFlush value.

Philippe: I have in theory a solution for the list of cleanups: make gROOTMutex a read-write lock.

Philippe: Due to CERN holiday let’s skip next week’s meeting.