I ran a series of timing tests of alala with a randome Essence image. I was looking at the variation of the different psphot steps with the total number of objects. The question of interest here is how our throughput compares to what we may expect from PS1.
subtracted background model: 2.776483 sec 567 peaks: 2.350294 sec 567 moments: 0.784404 sec identified 216 blended objects (0.006048 sec) rough classification (0.007589 sec) select psf model: 1.112221 sec built models: 0.005211 (567 objects) measure ensemble of PSFs: 1.903073 sec fit models: 5.739444 sec for 365 objects (77 psf, 33 ext, 255 failed, 202 skipped) replace models: 0.131781 sec (567 objects) measure ensemble of PSFs: 2.470944 sec measure full-frame aperture residual: 0.763284 sec measure magnitudes : 0.800974 sec for 567 objects (168 with apertures) replace background flux : 0.226746 sec
complete psphot run: 22.276232 sec extrapolation to GPC OTA: 66 sec
subtracted background model: 2.769444 sec 1069 peaks: 2.345380 sec 1069 moments: 1.478224 sec identified 515 blended objects (0.011470 sec) rough classification (0.008233 sec) select psf model: 1.326601 sec built models: 0.006923 (1069 objects) measure ensemble of PSFs: 2.135421 sec fit models: 7.566109 sec for 559 objects (206 psf, 51 ext, 302 failed, 510 skipped) replace models: 0.219286 sec (1069 objects) measure ensemble of PSFs: 2.849520 sec measure full-frame aperture residual: 0.942940 sec measure magnitudes : 1.222162 sec for 1069 objects (338 with apertures) replace background flux : 0.233185 sec
complete psphot run: 24.878902 sec extrapolation to GPC OTA: 75 sec
subtracted background model: 2.766679 sec 2050 peaks: 2.326407 sec 2005 moments: 2.521074 sec identified 519 blended objects (0.023435 sec) rough classification (0.010083 sec) select psf model: 1.410269 sec built models: 0.011972 (2005 objects) measure ensemble of PSFs: 2.791696 sec fit models: 10.465931 sec for 1114 objects (500 psf, 55 ext, 559 failed, 891 skipped) replace models: 0.437846 sec (2005 objects) measure ensemble of PSFs: 3.693040 sec measure full-frame aperture residual: 1.101108 sec measure magnitudes : 1.440022 sec for 2005 objects (478 with apertures) replace background flux : 0.226817 sec
complete psphot run: 30.939188 sec extrapolation to GPC OTA: 93 sec
subtracted background model: 2.746362 sec 3331 peaks: 2.344337 sec 3132 moments: 4.294272 sec identified 518 blended objects (0.036721 sec) rough classification (0.012682 sec) select psf model: 1.474402 sec built models: 0.019773 (3132 objects) measure ensemble of PSFs: 3.942925 sec fit models: 10.709155 sec for 1225 objects (617 psf, 55 ext, 553 failed, 1907 skipped) replace models: 0.846849 sec (3132 objects) measure ensemble of PSFs: 5.232133 sec measure full-frame aperture residual: 1.074216 sec measure magnitudes : 1.415196 sec for 3132 objects (471 with apertures) replace background flux : 0.227651 sec
complete psphot run: 36.642233 sec
subtracted background model: 2.743184 sec 26248 peaks: 2.483467 sec 23121 moments: 36.205507 sec identified 519 blended objects (1.431448 sec) rough classification (0.048091 sec) select psf model: 1.322417 sec built models: 0.086170 (23121 objects) measure ensemble of PSFs: 13.673267 sec fit models: 11.082951 sec for 1271 objects (583 psf, 63 ext, 625 failed, 21850 skipped) replace models: 4.120838 sec (23121 objects) measure ensemble of PSFs: 17.712378 sec measure full-frame aperture residual: 0.994449 sec measure magnitudes : 1.465164 sec for 23121 objects (435 with apertures) replace background flux : 0.234465 sec complete psphot run: 96.495855 sec
major stages
background 2.8 2.8 2.8 2.7 2.7 (Npixels) peaks 2.4 2.3 2.3 2.3 2.5 (Npixels) moments 0.8 1.5 2.5 4.3 36.2 (not S/N limited) choose psf 1.1 1.3 1.4 1.5 1.3 (Nobjects, S/N limited) ensemble.1 1.9 2.1 2.8 3.9 13.7 (Nobjects) fit 5.7 7.6 10.5 10.7 11.1 (Nobjects, S/N limited) ensemble.2 2.5 2.8 3.7 5.2 17.7 (Nobjects) apresid 0.8 0.9 1.1 1.1 1.0 (Nobjects, S/N limited) mags 0.8 1.2 1.4 1.4 1.5 (Nobjects, partial S/N limit)
Npeaks: 567 1069 2050 3331 26248 total: 22 25 31 37 96
insignificant
deblend (~ 0.02s) classify (~ 0.01s) guess (~ 0.01s) replace (~ 0.44s)
fitting dt vs n: moments: 1.4ms/obj ensemble: 0.5ms/obj full fit: 5.8ms/obj
Mark Pitts' analysis of 2MASS shows that we expect on average ~1500 sigma>25 sources and ~6200 sigma>5 sources per GPC chip, if we include the 107 deg-2 saturation limit. Note that If I limit the moments analysis to the brighter sources (all that is really needed), I can predict the following timing for PS1 (scaling Npix by a factor of 3):
back: 8.1 peak: 7.5 mmnt: 2.1 psf: 1.5 ens1: 3.1 fit: 8.7 ens2: 3.5 rest: 3.0 I/O: 9.0
total: 46.5sec (one cpu)
With our initial cluster of 16 boxes (64 cpus), we will be on the hairy edge of keeping up with the data if we are running at the full rate. However, by the time the survey starts, we should have two sets of boxes, giving us about a factor of 2 headroom. Though note that I'm ignoring any of the other steps of phase 2, not to mention any of the other survey modes. On the other hand, I'm also ignoring night-time, weather, longer exposures, and any other overheads. Still, it sounds like we will be tight for the first part of the survey until we get the third set of boxes (~Apr 2008).
