...
A comparison has been made between Pass7 and Pass8 reconstruction based on a single run (the 2nd run in the Fermi science data). There are three primary findings: timing, recon file size, and memory usage.
Details
P202 GR | 17-35-24 | |
P300 GR | 20-08-12 | |
Run | 239559565 | 2008-08-04 16:19:24 UTC |
#Evts | 2010995 |
...
2,010,995 |
Reconstruction processing time (Gleam)
Spreadsheet of job execution details:
https://docs.google.com/spreadsheet/ccc?key=0AveNVqKMtmVEdHJfbUNLMGs0S0NBcV9iaVhvMl9mdHc&usp=sharing
P202 (events/s) | P300 (events/s) | P202\:P300 ratio |
---|---|---|
6.71 | 2.56 45 | 2.62 74 |
RECON File size
P202 (GBkB/event) | P300 (GBkB/event) | P300\:P202 |
---|---|---|
|
|
|
...
14.67 | 37.27 | 2.54 |
Also see spreadsheet for exact numerology.
Memory footprint
Max memory reported by LSF
P202 (MB) | P300 (MB) |
---|---|
1800 (after 100,000 evts) | 1730 (after 20,000 evts – but with memory leak) |
Notes:
- P202 max mem after processing ~96,000 events
- P300 max mem after processing 20only ~20,000 events
- P300 starts at ~600MB and has a mem leak of ~55kB/event, which limits clump size. At ~60k-80k events, memory consumption has grown to make crashes inevitable, e.g.,
The exact time of a crash is unpredictable as it depends on the total memory resources and usage on a particular machine.Code Block terminate called after throwing an instance of 'std::bad_alloc' what(): St9bad_alloc sh: line 1: 25973 Aborted Gleam