using CPU % to adjust multiprocessor preferences

[JStateson]

OK - I read where you stated here quote "So 10% means the CPU ran 10% of it's time to feed the GPU"

Looking at the performance of an ATI 5850 on a Q6700, I see that collatz is a CPU hog and takes a full CPU.  As shown in the graph, the ati13ati collatz is highlighted in red (99.72) while it is working its way up and one that finished is shown at the top (99.52).  Note that the CUDA collatz "2.02 collatz" is NOT a cpu hog and used only 1.71% of a cpu.  Are you sure this is correct interpretation?



If so, then I should set the number of cpu's available on my quad from 4 to 3.  If I leave it at 4 then the the 4 QuantumFIRE alpha all take longer to finish.  However, this same system runs the ati13ati version of milkyway which uses less than %8 (7.42 as shown below).  This puts me in a bind because I dont need to set aside a cpu for milkyway.




Unless there is something I misunderstand, I am going to post over at collatz that their ati version uses almost a full cpu.  It does run extremely fast, and generates a lot of credits, but takes 4 minutes longer than a milkyway for about the same credits and burns a full cpu.


[EDIT} - I just checked the task manager and I do NOT see anything like %99 utilization for collatz.

[wicked]

[EDIT} - I just checked the task manager and I do NOT see anything like %99 utilization for collatz.

If you have a quad core CPU then 25% in Task Manager would mean the process takes one CPU (or actually one core). Seeing 99% would mean the process is using ALL four cores.

[JStateson]

The 99% was the value reported by boinctasks.

There is a discussion of this over at collatz where the developer states that the value for "time" returned by the ATI task is really GPU time and not CPU time.  This generates the %99 figure.  IMHO they need to change their code to report CPU time which is about 2 seconds for the entire work unit.  Milkyway correctly returns CPU time which then makes the boinctasks cpu% calculation correct, again, IMHO.

[fred]

The problem is with the application reporting back the wall clock time or something close to that to circumvent a BOINC problem. The time the GPU is running is of course always close to 100%.
What it should report back is the actual CPU time the feeder is using.