1) Message boards : Number crunching : fubar host of the day (Message 39050)
Posted 5 Jun 2019 by MaBoinc
Post:
Maybe we all should switch back to VirtualBox 5.2.16 and run single core tasks


If only that was just the solution, I've spend some time trying to figure it out but I honestly don't have a clue what is going on. Some observations though:

This host runs VB 6.0.8 and gets 23k for one task: https://lhcathome.cern.ch/lhcathome/result.php?resultid=230825272

I kinda hoped it would indeed be a VB thing, so I installed 6.0.8 and tried, still getting only 500-700 credits per task with similar run times on a very similar system (1950x instead of a 2950x):
https://lhcathome.cern.ch/lhcathome/result.php?resultid=230936322

Given the similar run times there shouldn't be that much of a difference, not consistently at least. The only real difference is the measured Device peak FLOPS, explained here: https://boinc.berkeley.edu/trac/wiki/CreditNew

This has been the case since LHC switched credit systems, explained by Laurence here: https://lhcathome.cern.ch/lhcathome/forum_thread.php?id=5021&postid=38790#38790

And indeed, 680.34/4.51*152.78=23047, so the difference in device peak flops indeed accounts for the difference. While the old system was far from perfect, this system seems to be even worse given how inconsistent it is.
2) Message boards : Theory Application : strange discrepency in credit points for Theory (Message 38981)
Posted 26 May 2019 by MaBoinc
Post:
From what I can tell it seems to be due to the large differences in device peak flops (the one which you can find above every task). Like this task https://lhcathome.cern.ch/lhcathome/result.php?resultid=230370834 from Vortac has a value of 152.78 GFLOPS on a 2950x, while I'm running a 1950x at 3Ghz @±37degC with quad channel 3000c15 memory, yet somehow this task: https://lhcathome.cern.ch/lhcathome/result.php?resultid=230662993 gives a value of 3.42 GFLOPS. Sure he might not have underclocked his Threadripper and it could very well be running on faster memory, but this seems a bit too extreme. Mine also used to perform perfectly fine before the credit update.
I updated virtualbox to the same version he is running but that probably won't make a difference, also ran the whetstone benchmark again, but not sure that will help either. Will report back if anything changes. Hope it can be fixed to be slightly more reflective of actual device performance.



©2019 CERN