ROOT I/O Meeting

Europe/Zurich
CERN

CERN

Brian Paul Bockelman (University of Nebraska-Lincoln (US))

Attendees:  Guilherme, Danilo, Viktor, Oksana, Jim, Philippe, [Zhe].

Viktor: Started as a fellow at CERN, my project will touch on ROOT I/O not quite yet clear how.

Philippe: There are upcoming non-forward compatible changes to ROOT I/O.

Viktor:  Yes, I have noticed and updating the JAVA version of ROOT I/O should not be too hard.

Viktor: A cluster is being constructed in Dusseldorf.  Will need to provide use case from CMS to drive the design of the cluster.

Oksana: Away this week.  I am investigating why the file merger is failing on a couple of the fast build.  It seems to be that the clang fast build given different result from other fast build.  Maybe it is that the flags is not properly seen.  Maybe it is missing some of the flags.  There is also some problems on the macs.  I would like to talk to Brian to profiling and see if the fast math really has benefit and if it does how much.

Jim: I am trying to get the bulk I/O pull request.  High level interface for numpy, but only after the PR is merge so that I can make direct PR.

Philippe: I am working on improving the performance of recursive remove (by re-allowing the use of the hash value and hence reducing from O(N^2) to O(N) … and then I will go back on to work on the lock itself.   I will also merger Brian’s request.

Danilo:  Working with Guilherme to address Dan Riley’s comments on Snapshots/ParallelMerge.  We still need his feedback to reproduce the excessive slow down.  

Jim: When I observed a slowdown of reading over-time, I discovered this was due to holding on forever to the basket and hence an ever growing memory use.

Guilherme:  How to add new option to the Snapshots actions (to reduce memory use, select compression algorithms, etc …).

Guilherme: Also adding the callback requested by CMS, this should be a small change.  One of the major request was to be non-blocking.  Whenever a block is removed from the queue, the callback(s) are executed.   We need to get feedback from CMS if this match their need.

Guilherme: The merger will count the number of clients and as soon as we remove the same number of block as client, we notify.  We could also do something per client.

Philippe: The ROOT I/O workshop is coming up.
We should have an update on the parallel buffer merge and snapshot interface.

Guilherme: I will prepare such a talk and add more comprehensive performance numbers.

Philippe: Need JIRA ticket for Jakob’s discoveries (mapping of objects/pointers not needed).

Philippe: Jim can you present your LZ4 numbers.

Jim: will do.

Oksana; Working on adding the new zlib version.  After that I will need to make a presentation on ROOT modernization.  Need better.

Philippe: Also need back port of lz4 build system update. and test.

 

There are minutes attached to this event. Show them.
    • 16:00 16:20
      Round Table 20m