Choose timezone
Your profile timezone:
ROOT I/O Meeting
Attendees: Jim, Oksana, Jakob, Philippe, Axel.
Jim: RAS (awkward array).
Oksana: Away from the week. Now working on tutorials for Software Foundry. Thinking about the ROOT developer retreat.
Philippe: Debugged problem that was due to a missing dictionary which meant that the old schema StreamerInfo was not being attached to the TClass …
Jakob: Pull request from Simon on visualization of RNtuple files that let you see where in the file each column data is.
Jakob: Next week RTuple workshop between ROOT team and CERN IT department’s storage group/scientific-computing group. Goal is to discuss together and see where RNTuple can be tune to better use the storage system and vice et versa.
Jakob: Worked file format issues, in particular adding RTuple inside a ROOT file. RTuple itself is a directory and Data are stored in TKeys inside that directory (which is bad because of the directory’s hash list). So now importing a the scheme a TTree which a Meta-Data object that is a key and that object has the location on file of the header and footer which have the location of the anonymous keys (not registered anywhere) that contains the data. We will then need to upgrade hadd/RNTuple merge to properly update the meta data.
Jakob: to speed-up concatenation, wouldn’t it be cool to have a file-system level way to concatenate 2 files by having the file system simply update its meta data.
Philippe: Yes, but unless the file system implement COW then the 2 original files would cease to exist, isn’t it?
Jakob: Maybe. Maybe EOS can implemented this?
Jakob: Continuing on file format issue.
Jakob: Some of the work is already document in my branch/fork.
Axel: 2 things to bring up. (1) Simon and Jakob have implement Float16 and the like. At the C++ meeting, that you could have zero-copy *AND* float16 by zero-ing the bits rather than bit-packing. However I do not know the performance trade-offs.
Jakob: Was there some experience with something like that?
Jim: Yes within CMS, Andrea Rizzi. The result was not stellar with LZ4.
Jim: The current scheme is much better than zero-ing out as it can encompass both the precision reduction and the range.
Philippe/Axel: I/we still think that it is worth measuring in the context of RTuple.
Axel: Grid Deployment Board meeting, I ask them if they could increase the minimum cpu on the Grid to AVX2 (which would be the first time that we make an architectural change since the 32bits to 64bits). Do you remember the cloud fare performance difference.
Oksana: Compression: twice as fast. Decompression: similar.
Axel: Was CMSSW ever tested with AVX2?
Oksana: I am not sure, I will ask.
Axel: At the ROOT planning meeting, there was wide interested in testing ZSTD.
Oksana: ATLAS does not yet have any file that can be read by bare ROOT.
There are minutes attached to this event.
Show them.