[parsec-users] sub-linear scaling

cbienia at CS.Princeton.EDU cbienia at CS.Princeton.EDU
Mon Jun 23 00:21:37 EDT 2008


Hello Major,

Sorry for the late reply, I'm currently at ISCA in Beijing and have 
only limited email access.

Machine constraints can limit the speedups that you can achieve on real 
machines. This is true for all parallel programs, not only PARSEC 
workloads. Especially limited memory bandwidth is an issue for CMPs 
like the AMD Quadcore which you used. The speedup results which we 
report in the paper are for an ideal architecture. The point was to 
demonstrate that sublinear speedups are unlikely to be an issue of the 
workload (it's not a full proof because we only report upper bounds, 
lower bounds would be more useful but are complicated for the reasons 
we describe in the report).

- Chris



Quoting Major B <mbb45 at cornell.edu>:

> Hi,
>
> I find several PARSEC benchmarks to scale sub-linearly on quadcore 
> machines (AMD Quadcore Phenoms and Intel Quadcore Q6600) using native 
> input sets.
> I'm running with several GBs of DDR2-800 RAM off the local hard 
> drive. I tried over NFS but that only gave worse results.
>
> Has anyone else encountered this or found all benchmarks scale at 
> least linearly to four threads? I noticed the technical report did 
> not use native inputs. Are there any other sources that have?
>
> Thanks!
>
> -Major
> _______________________________________________
> parsec-users mailing list
> parsec-users at lists.cs.princeton.edu
> https://lists.cs.princeton.edu/mailman/listinfo/parsec-users
>





More information about the parsec-users mailing list