[parsec-users] Scalability of PARSEC benchmarks ?
cbienia at CS.Princeton.EDU
Tue Apr 1 17:26:36 EDT 2008
Exactly, you should get very high speedups in almost all cases. Only bodytrack
has a known limitation which we're working on. If you run into any problems
I'd be interested to hear about it. AFAIK nobody has tried to run experiments
with 128 cores yet.
You might also be interested in the section discussing the upper bounds on
speedups which we have in the tech report.
On Tuesday 01 April 2008 03:02 pm, Roberto Giorgi wrote:
> Hi Chris,
> ... so you mean that if I have perfect memory and perfect
> interconnection network,
> then there is no bottlneck in the applications (or at least the serial
> portions are minimal) ?
> Hello Roberto,
> How well the PARSEC workloads scale in practice depends on the machine. All
> programs scale well to a large number of cores if their requirements are
> met. In practice (or if you simulate a realistic machine) you will
> frequently hit memory bandwidth limitations soon.
> If you only care about good scalability start with programs that have small
> working sets first. Depending on your architecture intense communication
> between threads can consume a lot of memory bandwidth, too. The commercial
> workloads which use the pipeline model tend to be tougher on the machine in
> that sense.
> The more your workloads have those two properties (large working sets &
> intense communication) the more challenging it will be to achieve good
> speedups. Have a look at the characterization to get an impression of the
> requirements of PARSEC programs.
> - Chris
> On Tuesday 01 April 2008 07:03 am, Roberto Giorgi wrote:
> > Hi, I'm new to PARSEC: can someone please tell me if these benchmarks
> > can be used in scalbility studies, e.g. at least up to 64-128 cores (like
> > SPLASH-2)?
> > In case of positive answer: which benchmark could I first start with
> > (simpler, more scalable with number of cores)?
> > Thanks!
> > Roberto.
> parsec-users mailing list
> parsec-users at lists.cs.princeton.edu
More information about the parsec-users