sPHENIX tracking discussion

[pinkenburg <pinkenburg AT bnl.gov>]

Hi folks,

Tim put his multi threaded emc pulse fitting into the prototype 
reconstruction and I have been running tests on our idling test node 
(spool1150, 96 cores).

Initially it looked like there is a nice gain but then I found that some 
input files didn't exist or only had a few number of events so we had 
idling cores on that node. Once everything was sorted out and 96 jobs 
ran over 100k events there was no improvement comparing single threaded 
with 3 threads (which according to Tims studies gave the best 
performance). Actually there seems to be some indication that running 
multithreaded took slightly longer.

These tests were done running over local files (from an ssd), reading 
directly from luster still needs to be done. The rate is not that high - 
I don't think there will be a different outcome.

It looks like there is basically no free lunch for us, the huge 
performance increase we saw on the shared pool are likely due to the 
availability of idling cores due to i/o limited STAR and PHENIX jobs 
(whereby STARs jobs often blocked any job using afs from running at more 
than 30%).

I would need an example macro to run the multithreaded tpc clustering on 
spool1150 but given this (with the results from the sim production), it 
seems the gains we saw from multithreading just evaporated.

Chris

--
*************************************************************

Christopher H. Pinkenburg       ;    pinkenburg AT bnl.gov
                                ;    http://www.phenix.bnl.gov/~pinkenbu

Brookhaven National Laboratory  ;    phone: (631) 344-5692
Physics Department Bldg 510 C   ;    fax:   (631) 344-3253
Upton, NY 11973-5000

*************************************************************