Round table
Vincenzo
- Resumed work on making the RDataFrame constructor accepting a file name infer the data format inside the file, so that TTree and RNTuple based analyses can use exactly the same code. The PR is practically ready to be merged
Serhan
- Prototype for RNTupleMerger is merged, a few missing items discussed on the PRs and will be followed-up
- Make sure that the merger works with deferred columns, projected fields, alias columns
- RNTuples with a fixed schema already work on the current prototype, tested on real ATLAS DAOD files.
- RNTuple in ATLAS, we want infrastructure support for the larger scale tests next year. We have code that expects TTree in the files (mostly concerning job configuration). Updating them so we can run production jobs with RNTuple.
- Should be within a few days reach.
Jonas
- Progress on creating RNTuple parallel writer prototype. Every thread starts preparing a cluster and there is only very lightweight locking to commit a cluster to disk.
- Prototype available, scales ok-ish. The writing is still under a lock, this will be removed. Already now there is compression happening in parallel. Able to fully saturate IO bandwidth on spinning disk on
ntpl-perf01
, we are off by a factor 2 for the SSD bandwidth.
There are minutes attached to this event.
Show them.