Done by TJ Lane and cpo on Oct. 31, 2019
1M XTC Events from "xtcwriter -n 1000000 -e 0" on the test stand ffb (/ffb01/cpo/)
10,000 batch size (both for smalldata and DataSource)
10,000 cache size
"n" is the total number of cores given to mpirun on drp-tst-dev003 (a 20 core node)
Conclusions
- small h5 production is a 10% perturbation on the event-loop time, and we can scale our way out if performance is an issue.
- hdf5's VDS (done by "smd.join_files()") seems to behave well with 500 files each with 1M events, which is good because that is the one thing we can't scale.
Measure time for run.events() and smd.join_files()
=== time (sec) ==
loop_time join_time
--------- ---------
SRV=0 n=16 [no smalldata]
10.3 --
11.7 --
10.6 --
10.3 --
10.3 --
10.4 --
SRV=1 n=17
11.7 .0063
11.5 .0066
11.6 .0066
11.4 .0067
12.1 .0061
11.6 .0065
SRV=2 n=18
11.5 .0072
11.3 .0076
11.3 .0074
SRV=4 n=20
11.7 .0105
11.4 .0105
11.2 .0103
|
Measure smd.join_files() for many files
join times (seconds):
each file was 1M data points
#files time
1 0.0065
2 0.0075
3 0.0104
100 0.5900
500 1.2000