Details
-
Type:
Story
-
Status: Done
-
Priority:
Major
-
Resolution: Done
-
Fix Version/s: None
-
Component/s: afw
-
Labels:None
-
Templates:customfield_11800 17727
-
Story Points:5
-
Epic Link:
-
Sprint:Science Pipelines DM-S15-3
-
Team:Data Release Production
Description
In DM-1128 we implemented span-based dilation for footprints. A brief test on synthetic data indicated that this was a performance win over the previous version of the code.
In May 2015, this code was merged to HSC and applied to significant quantities of real data for the first time. A major performance regression was identified:
[May-9 00:26] Paul Price: processCcd is now crazy slow.
[May-9 00:29] Paul Price: Profiling...
[May-9 00:40] Paul Price: I'm thinking it's the Footprint grow code...
[May-9 00:44] Paul Price: And the winner is…. Footprint construction:
[May-9 00:44] Paul Price: 2 0.000 0.000 702.280 351.140 /home/astro/hsc/products/Linux64/meas_algorithms/HSC-3.8.0/python/lsst/meas/algorithms/detection.py:191(makeSourceCatalog) 2 0.005 0.002 702.274 351.137 /home/astro/hsc/products/Linux64/meas_algorithms/HSC-3.8.0/python/lsst/meas/algorithms/detection.py:228(detectFootprints) 15 0.001 0.000 698.597 46.573 /home/pprice/hsc/afw/python/lsst/afw/detection/detectionLib.py:3448(_init_) 15 698.596 46.573 698.596 46.573 {_detectionLib.new_FootprintSet}
[May-9 00:53] Paul Price: If I revert HSC-1243 ("Port better Footprint-grow code from LSST"), then the performance regression goes away. @jbosch @jds may be interested...
The source of the regression must be identified and resolved for both HSC and LSST.