Skip to content

Mitigate the large NSW segment maker time

Hi everybody,

the CPU footprint of the muon reconstruction is unbearable high this year. The cause is a very talkative NSW. In particular, the micromegas are telling very exciting fairy tales.

image

In !62170 (merged), @pscholer cut off the super communicative events with processing times per event beyond 30 minutes. Yesterday, @pscholer, @goblirsc and myself were sitting together and had a more thorough look at it. We came up with 2 ideas which are implemented in this merge request

  • Detection of burst regions*: Before the hits are passed to the micromega seeding, the cluster channels are binned into a histogram. The bin width is for the moment set to 100. If a bin contains more than 5 clusters, all clusters in this bin are removed. In the next step, two neighbouring bins are summed up. If a pair contains more than 7 clusters, both bins are emptied as well. Both numbers are initial guesses and probably require some fine-tuning.

  • The position of the micromega seed in the transversal plane should roughly match the transversal seed direction. A first test on 1000 data events indicates very promissing results:

Vanilla:

00:15:31 PerfMonMTSvc                                                      INFO Execute        999       1560690.38     761808         264259              MuonSegmentMaker

Patched:

01:08:14 PerfMonMTSvc                                                      INFO Execute        999       105792.21      0              227655              MuonSegmentMaker

A first validation on a particle gun di-muon sample nicely demonstrates that the patch does not have an impact on the inner hit distributions.... image

image

image

However, this is an ideal unicave scenario and we are expecting that the segment making becomes sensitive to pile-up. To determine the optimal setting of cuts we need to roast a few polar bears and reprocess a high-pileup with the patch and maybe also with a few variations of the window size

image

I guess, the most easiest approach is using the grid. So, I'm humbly asking for an increase of the grid quota for the det-muon account

Tagging: @nstyles, @jcatmore, @christos, @pscholer, @goblirsc, @jmaurer, @strichte, @amete, @zmarshal, @girolamo, @nicolaid, @mvanadia, @dhayden, @stavrop, @iodice, @sabidi, @atlasbot, @sangelid, @pgadow, @sroe, @syan

image

Full list of particle gun plots here

Edited by Johannes Junggeburth

Merge request reports