Topic: HEP-CCE2/SOP video conference
Join Zoom Meeting
https://cern.zoom.us/j/63214151964?pwd=OXNtZWJuZFBvT3FWN1RhWG9EMzE1QT09
Meeting ID: 632 1415 1964
Passcode: 528737
# Attendees
- Peter, RobR, Dan, Doug, RobL, Rui, ChrisJ , Amit, Marcin, Matti, Paolo, Shane, Suren, Jean Luca
# Introduction/Management News
Saba joining: slight modifications to agenda; leads
This collaborative note – everyone can pitch in
Root RNTuple meeting in two weeks: model field and entry
Jean Luca LBNL joining us for a few months: introductory talk:
We have some tools for collecting information from the i/o stack layer, but less clear how to turn that into tuning actions
Drishti: can consume Darshan, Recorder, HDF5 Trace information
Recently Able to extract HDF5 related traces and compare with Darshan DXT
Produces many plots visualizing i/o behavior
Finds issues, warnings and recommendations to improve performance of the code
Traces issues not only in user code but also in application (was able to find some MPI Write related issue in HDF5)
PC: In HEP community, ROOT is more common than HDF5, how well or how would it work with other applications other than HDF5 (like ROOT)?
Can trace POSIX or STD IO
If Darshan provides DXT from ROOT, DRISHTI can be used for instrumentation
Recommendation from Drishti is more like one problem to many possible solutions (could trace problems at various layers eg from HDF5 to underlying MPI-IO)
# Applying Lessons Learned to HEP Experiments, Mimicking Framework, Darshan and HDF5
Mimicking Framework, Darshan and HDF5
Amit Bashyal, Christopher Jones, Kenneth Herner, Patrick Gartung , Peter van Gemmeren, Rui Wang, Saba Sehrish, Shane Snyder
Rob L: Test framework darshan log excluding everything and including only RNTuple related.
Chris: Input source that Rob is using is very very simple and hence reflection in the darshan log.
# Tracking and aiding the evolution of ROOT I/O, in particular RNTuple
Alaettin Serhan Mete, Amit Bashyal, Barnali Chowdhury, Christopher Jones, Daniel Riley, Kyle Knoepfel, Marcin Nowak, Matti Kortelainen, Peter van Gemmeren, Philippe Canal
CJ: Slides uploaded for Performance Measurement for RNTuple using mimicking framework
Throughput and memory usage comparison between TTree and RNTuple
Hacked MiniAOD data so that RNTuple could read certain classes of data that RNTuple could not read. Data that cannot be copied to new RNTuple ~10%
Both TTree and RNTuple gets data from hacked data
Comparing RNTuple with 3 different configuration of TTree (unsplit, fully split and mixed split of TBranches)
Mixed is: Mixing based on which branch is better to split and which is better to not split
Single threaded concurrent job restricted to 12 cores (in case of out of memory for RNTuple related test for more than 12).
RNTuple does very well compared to TTree
RNTuple scales perfectly upto 8 threads.
ParallelRNTupleOutputer does not uses IMT. Communicated with ROOT developers.
Uses lots of memory, limits running jobs upto 4 threads and restriction on concurrency to avoid memory swap.
RNTupleOutputer with IMT works as good as ParallelRNTupleOutputer
RNTuple requests 10 times more memory than TTree (resulting in overallocation, CJ’s take and passed information to RNTuple Developer and now working on this to bring down the memory requirement)
Dan: Might be interesting to use memory allocation request vs. compression algorithm used.
Chris: TTree is also using LZMA like RNTuple.
Peter: Some of the memory usage might depend upon data model etc but the current request is still something to address.
AB: Ongoing clean up of CAF RNTuple Writer to integrate into CAF Maker package of DUNE.
Next Review meeting
# Reduced Precision and Intelligent Domain-specific Compression Algorithms
Amit Bashyal, Meghna Bhattacharya, Peter van Gemmeren
AB: Integrated IDEAL-EM in the compression test framework. Tests done with both fake data and waveform data from Majarona experiment to see the performance. The compression parameters are stored as JSON string but complete integration going on. Next step is to write automated exploration of compression parameters to achieve certain compression ratio for each algorithms.
Will present next week.
# Object Stores and Strategies for Data Placement and Replication
Bo Jayatilaka, Doug Benjamin, Nicholas Smith, Rob Latham, Rob Ross, Saba Sehrish, Shane Snyder
DougB: Got permission from owner of EIC s3 server to use it for RNtuple testing at BNL. (of course I will need to look at the RNtuple core code to see if Amazon S3 code to see if code changes are needed)
# Optimized Data Delivery to HPC systems
Marco Mambelli, Saba Sehrish
Nothing here.
# Any Other Business