Just a 1-step setting (like the current approach), or would the idea be dynamically be scaled further? I.e. if the non-BOINC system apps continually use 50% for some time, BOINC limit will be lowered to 2 CPUs, after rising above 75% just 1 CPU, etc.
Peter Slacik David Anderson wrote: > The suggestion is: > if the non-BOINC CPU load on a 4-CPU system is 25%, > BOINC should use only 3 CPUs. > Seems reasonable to me. > Comments? > -- David > > -------- Original Message -------- > Subject: RE: Enhancement suggestion: CPU Usage exceeds parameter > Date: Tue, 3 Aug 2010 20:57:43 -0400 > From: Sean White <[email protected]> > To: 'David Anderson' <[email protected]> > > I though NCPU's was not the % of CPU's, but the number BIONC was allowed to > use? In the parameters panel - this is listed as % of processors - so if > this is the units, then your expression makes sense. How would you mange > the current 'over allocation' allowed which allows you to fully task the > CPU's as well as GPU's? Or -- if the over allocation is A, then BIONC would > allow for peak processor usage to spike to NCPU's-X+A, and continue to push? > > > The reason for a more involved process would be to attempt to detect when > the user or another pre-scheduled task is actually using the computer (i.e. > gaming, watching a video, video recording etc) where the user experience or > task is often time sensitive and may peak to higher usage levels. The more > complicated approach assumes that the only tasks running on the computer > that would consume any significant fraction of a single CPU's time are > 'critical' or 'important' tasks. Assuming that there is some element of > time-criticality to any of these means that BIONC would need to back off > further than just "x" to give sufficent room as to minimize the time > critical impact. I suppose that this could be done by monitoring both the > average non-BIONC CPU usage over a period as well as the Standard deviation > of the CPU usage. If the average is over a threshold (configurable?) then > BIONC scales back by X + 2 Standard Deviations to ensure that overhead is > available proportional to the peak processor usage over a period. If the > average usage is below the threshold, the 'noise' is ignored. > > -Sean. > > > > > -----Original Message----- > From: David Anderson [mailto:[email protected]] > Sent: Monday, August 02, 2010 7:09 PM > To: Sean White > Cc: BOINC Developers Mailing List > Subject: Re: Enhancement suggestion: CPU Usage exceeds parameter > > That's a good idea. > I'd been thinking about something similar. > What I was thinking was to eliminate the preference, > and just change things so that if non-BOINC CPU usage is X, > BOINC will use at most NCPUs - X. > -- David > > On 26-Jul-2010 2:37 PM, Sean White wrote: > >> David, >> >> >> I've been a longtime BIONC fan, and had an 'enhancement' suggestion with >> respect to the 'cpu usuage' exceeds 'parameter' feature. Instead of >> shutting down the entire set of processess when the cpu usage exceeds a >> threshold, it would make more sense to roll back usage 1 CPU at a time. >> >> If a computer has Y CPU's available for use, and the 'other process >> usage' parameter is set to value X, then when Non-BIONC Usage /Y > X for >> a minimum time interval 'T', then we freeze one BIONC process, freeing 1 >> CPU to manage the non-BIONC tasks. The next trip occurs when Non-BIONC >> usage / Y > (1/Y+X). (I.e. when the non-BIONC tasks load up the 1 CPU >> and start impinging on the 2^nd CPU, we drop the second CPU out). I >> would suggest that a reasonable 'drop out time' is 1s or so? (whatever >> is currently used). >> >> This approach would eliminate the annoying 'entire BIONC all tasks >> offline' event which I encounter when multiple tasks happen to end at >> the same time, or when you open another application that happens to need >> all of 1 CPU briefly to get going. >> >> Cheers, >> >> >> Sean W. >> >> > _______________________________________________ boinc_dev mailing list [email protected] http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address.
