The run time is mainly determined by the number of pixels which are being annealed, so estimate the 90th percentile of the number of pixels to anneal. Pixels which have been annealed have confidence≥60, with those in which the annealing solution is returned having confidence=90. The following plot shows the number of pixels with confidence≥60 as a function of time for typically one image per day from 2011.07.01 through 2013.06.30. Times which do not have quality=0 are excluded. The x-axis is relative to 2011.01.01. The 90th percentile of confidence≥60 is at approximately 7.2×105 pixels.
The first speed tests have been done for 2011.02.01, which has approximately 4×105 pixels to anneal, 2012.06.01, which has approximately 5.8×105 pixels to anneal, and 2012.01.01, which has approximately 7×105 pixels to anneal. Thus, focus on run times of 23000s*4×105/7.2×105≈13000s for 2011.02.01, 19000s for 2012.06.01, and 23000s for 2012.01.01 to estimate the best annealing schedule. The following plots show the final energy from the annealing as a function of the run time, for a range of values of neq (different colors) and values of tfctr (different points in the same color, not labeled). Each point is the mean over a few random number seeds, with error bars given by the standard deviation.
In this case, the best annealing schedule for 2011.02.01 has neq≈20 and tfctr≈0.99, for 2012.06.01 it is neq≈50 and tfctr≈0.98, while for 2012.01.01, it is neq≈100 and tfctr≈0.95, with very large uncertainties due to the small number of random number seeds used so far. For 2011.02.01, this is essentially what is presently being used for NRT HARPs, and somewhat faster cooling (i.e., worse answer) than is presently being used for definitive HARPs (neq=100, tfctr=0.98). Thus, we could perhaps disambiguate full disk every 12 minutes with quality somewhere between the present NRT and definitive HARP disambiguation. However, there is a distinct difference in the optimal parameters between these two dates.
The second alternative for full disk disambiguation is to run once over 4-6hr using one 8-core node. Thus we have 8*4hr=32CPU-hr=115200CPU-s for each full disk disambiguation. This is slightly longer than the longest run shown above. How do the results compare? The left image shows the radial component of the field with contours of the confidence; the right image shows the difference in azimuth between neq=200, tfctr=0.98 (the slowest annealing schedule tried) and neq=100, tfctr=0.96 (a plausible annealing schedule for keeping with at 12 minute cadence). Note that areas outside the blue contour have the same disambiguation regardless of the annealing schedule.
To be quantitative about it, the fraction of pixels with confidence=90 with the same disamibguation between neq=200, tfctr=0.98 and neq=100, tfctr=0.96 is 0.986; for pixels with confidence≥60, the fraction is 0.982. The change in energy between these two annealing schedules is a result of changes in a small fraction of the pixels.
To really establish the annealing schedule, we should: