[parsec-users] influence of interleaving and undeterministic
Christian A Bienia
cbienia at CS.Princeton.EDU
Sat Sep 4 04:48:16 EDT 2010
I don't think you can ignore it because it's not clear which measurement is the 'correct' one. But you can do many runs and use statistical methods to compute the averages with standard deviations of each metric. As far as I know that is the only proper way to deal with the variability of parallel programs. Also see Alaa Alamaldeen et al. "Addressing Workload Variability in Architectural Simulations" about this approach for simulations.
----- Ursprüngliche Mail -----
Von: "lulu he" <loloseed at gmail.com>
An: parsec-users at lists.cs.princeton.edu
Gesendet: Samstag, 4. September 2010 03:28:45 GMT +01:00 Amsterdam/Berlin/Bern/Rom/Stockholm/Wien
Betreff: [parsec-users] influence of interleaving and undeterministic
I'am using pfmon to profile PARSEC.
Because my machine only has 2 PMU ,
in order to get cache miss rate every 1M insns , I have to collect 3
events(insns_retired , cache_reference , cache_miss).
So ,this is the problem, I have to run the program twice(once
insns_retired+cache_reference , twice insns_retired+cache_miss) to
collect 3 events.
sometimes, I got cache_miss_ratio>1.0 because of undeterministic. But
most time , results seem OK.
I want to ignore the influence of such undeterministc stuff.
parsec-users mailing list
parsec-users at lists.cs.princeton.edu
More information about the parsec-users