Choose timezone
Your profile timezone:
Powered by Indico
v3.3.3-pre
Attendees: Philippe, Danilo, Oksana, Jim, Chris Jones, Guilherme, Peter VG.
* CMS vs TBufferMerge, where are we?
* TTreeCache caching updates
* Upcoming ROOT I/O workshop.
* Block I/O (if Brian can make it)
* Cloudfare
Danilo: Met with CMS and agreed on a candle CMSSW example for the parallel buffer merge. We started a discussion on how to improve the interface to be more specific/useful for framework, including a non-blocking interface. Also need to have a way to avoid infinite accumulation.
Guilherme: The conversation was productive and I know understand better the CMS need. We need to update the code/interface to better match those needs. I also have access to a new example and will be example to extend our test based on that.
Chris: We discovered a mysterious scaling problem discovered yesterday likely involving implicit IMT. The other candle is a representation of what Dan R. has been doing.
Danilo: Shared_ptr is planned for v6.16, I need to work with Philippe on it.
Philippe: TTreeCache. The current TTreeCache implementation is vulnerable to over and under reads in many scenario that deviates from the usual (well clustered file, all entries read in order).
Chris J: We need a new tag for access to TBranch::BackFill in v6.10 and v6.12.
Axel: Indeed, next week.
Jakob: Integrating comments from the team meeting, updating the source code of the new TTree interface proposal.
Jakob: I recommend Jim’s presentation on I/O implementation comparisons done at the HSF at Napoli.
Jim: yes it includes a description of the properties various format/libraries and when you might want them or .. not.
Oksana: Cloudfare has 2 PRs with some more work to be done. I am investigating why even-though it is 50% it shows in some example a 2% increase in file size. Brian claims it might be related to the difference in hashing algorithm/function used for dictionary lookup. Looking at more example we see both increase and decrease in file size (from -2% to 2% and one outlier at 7%).
Brian: level 1 traditional only 15% faster than Cloudfare level 9. So the decrease in size can be made up by increase the compression level.
Oksana: I took away the file that was not used by ROOT (problem with CRC32 with assembly instruction that had a license not compatible with ROOT’s). I will rebase the PR and then they are ready.
Oksana: Will focus on Cling modulation. Then I might look into Bulk I/O. And after CHEP I would go back to Zstd.
Brian: We have 3 options for Bulk I/O: (a) wait for Brian to come back from Proposal writing mode (b) Have Oksana and Zhe to take a look at the PR and updates (c) any other ideas.
Axel: (c) may be ‘just’ write a presentation on interfaces and what it works.
Brian: If it is just description I should be able to make a presentation on Monday 23rd at the ROOT Team meeting.
Jim: I will be able to help with the PR update since it contains changes related to PyROOT.
Zhe: I submitted a pull request for the signal handler. I have written the shell script for lldb and it works in my local testing. (it is based on the gdb shell script).
ROOT I/O Workshop:
Danilo: I will contact ATLAS to find a speaker.
Chris: I will think about it for CMS.
Philippe: I will contact Peter about ALICE.
Brian: Bulk I/O presentation at the workshop. If I can’t make it to the workshop, I would come for a few days at FNAL instead.