ROOT I/O Meeting

Friday 25 May 2018, 16:00 → 17:40 Europe/Zurich

32/1-A24 (CERN)

Brian Paul Bockelman (University of Nebraska-Lincoln (US)), Philippe Canal (Fermi National Accelerator Lab. (US))

Attendees: Philippe, Peter, Jakob, Oksana.

Peter: Got a summer student working for Taylor and a staff at BNL looking at file merging using MPI.  The ATLAS I/O Review is still
in the planning stage, currently tentatively September 26, 27; nothing official yet.  (Atlas week 25 to 31 at DESY).

Jakob: While looking at the TTree::Fill, I found it brings ~3% performance improvement in the serialization of floats.  On my system, the file merger is failing, likely thanks to a newer version of LZ4 on my system.  I will check that this is ‘solely’ due to the better lz4 and increase the tolerance accordingly.    I have a couple more ideas to improve TTree::Fill.   

Guilherme: Working on Vectorization of GenVector which has been vectorization using Vc directly, I updating to use VecCore.

Zhe: I have been looking at LZ4 on the example on how to use dictionary to provide random access compression/decompression.  [When I tried with zlib, I had to do too much work to extract ‘fake’ out the dictionary functionality and ended-up with bad compression ration].  Since there is a lot of flexibility in how to deal with dictionary (size of the dictionary).  Two phases, generate the dictionary and then compress.  

Philippe: Yes, this is interesting and we should pursue it.

Zhe: See https://fossies.org/linux/lz4/examples/dictionaryRandomAccess.md.

Oksana:  I resolve the fact that the StreamerInfo record was (now) always compressed with LZ4 thus leading to the file not being unable to read the file with older version of ROOT even when the user requested ZLIB compression.  So now we compress this record with the same compression algorithm and level  as the rest of the file (i.e. allowing this record to not be compressed).   

Philippe: Yes, this can be merge after we check if roottest needs to be updated.

Philippe: I am investigating tp_process_imt, fixed a problem with the IMT lock but still see high variability of the download time (hence timeout).  We are trying to figure out what’s going on with root.cern.ch.

 

16:00 → 16:20 Round Table