Open olantwin opened 3 years ago
Locally, no individual part seems to be a bottleneck. Suspicion is that it's EOS/remote I/O. EOS very heavily caches, so up to two orders of magnitude faster for cached events.
Maybe events could be prefetched/reading moved to a separate thread to not block everything else?
Not sure, why exactly this is only an issue for the hdf5 reader. Maybe the format is "too smart" and underlying assumptions break down on EOS?
h5py is not faster, neither locally nor on lxplus/EOS