An asynchronous library for reading ROOT files over xrootd.
This package is an investigation into using asynchronous programming via async/await
syntax in python 3 to make remote reading of ROOT files as concurrent as possible.
The package takes a lot of inspiration from uproot.
The initial goal is to provide an equivalent to the uproot.numentries function, which is optimized
to extract the number of entries in a given TTree embedded in a ROOT file as quickly as possible.
In both implementations, the typical file requires 3 synchronous xrootd read() calls, however in
this library we: trigger the sequence of calls concurrently for all files to be read; do any decompression
in a thread pool outside the loop, and never block on read IO so the main thread can continue unpacking
the structures. The net effect is an on average 5x speedup for a list of 10 files as shown in demo.py.
Running this demo at home (in the US):
$ ./demo.py
Elapsed (async): 3.28 s
Elapsed (uproot): 24.47 s
All entries agree? True
and running it on a CERN lxplus machine (very close to the file servers):
$ ./demo.py
Elapsed (async): 0.05 s
Elapsed (uproot): 0.24 s
All entries agree? True
In preparing this implementation, the structure unpacking functions had to be largely reworked from uproot, as there the unpacking and IO are heavily intertwined. I think the sans-io philosophy may apply also to this case, and uproot could easily become both a sync/async library if the structure unpacking is separated from IO fully.