[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Condor-users] Why do kflops and mips vary so much?



> > Actually, there have been discussions here about whether it makes sense
> > to rerun the benchmarks at all.  After all, how often does a CPU get
> > slower or faster while the system is running?  Yes, I know that CPUs
> > these days *do* change their clocks, but nonetheless, these benchmarks
> > are indented to capture the system's capabilities, which should never
> > change.
> 
> it's an interesting idea, but as data centres move to an active power
> management model: shutting down cores, reducing clockspeed and voltage in
> response not only to internal demands but external ones too (to the minute
> power prices for example) pretending that they never change is perhaps not
> the best option.
> 
> In many cases they don't change much, or are configured to not do so, so
> big clusters might benefit from just running it at start up and reporting
> that value for the life of the instance. Deciding that should be the way
> forward for everything is liable to be problematic in future me thinks.

In that case, you'd like what just came out of a hallway discussion...  I 
suggested that the benchmarks should be run as a "hawkeye" module...  There 
should be a means of running a module just once at startup anyway.  Thus, your 
startd's hawkeye job list would contain the benchmarks you want to run, and 
how often to run them.  Additionally, this would give users the ability to add 
/ remove benchmarks that are / are not meaningful to them.

Thoughts?

-Nick

-- 
           <<< It's the question that drives us. >>>
 /`-_    Nicholas R. LeRoy               The Condor Project
{     }/ http://www.cs.wisc.edu/~nleroy  http://www.cs.wisc.edu/condor
 \    /  nleroy@xxxxxxxxxxx              The University of Wisconsin
 |_*_|   608-265-5761                    Department of Computer Sciences