[parsec-users] PARSEC indeterminism and Prefetching
Muhammad abid Mughal
mabidm_pieas at yahoo.com
Mon May 9 07:11:47 EDT 2011
Thank you very much for answering.
inst count differ = Inst count(base)-Inst count(sys. with prefetcher)
= 408629113 - 376282439
My base system has more inst count.I use gems-2.1 for simulation i.e.
i am going to collect trace from my base system and i will use trace fro
simulation.This way both system exactly has same amount of inst. count.
Thats the only solution to avoid indeterminism in PARSEC.
From: Mark Roth <mroth at sfu.ca>
To: PARSEC Users <parsec-users at lists.cs.princeton.edu>
Sent: Monday, May 9, 2011 13:39:54
Subject: Re: [parsec-users] PARSEC indeterminism and Prefetching
How much do the instruction counts differ? I think you would expect to
see some level of difference in instruction counts with different
system set ups depending on the amount of contention. Competing for
locks, spin waiting, etc. will all lead to an increase in the total
number of instructions retired.
Having the pre-fetcher turned off could be lowering the amount of
contention in the system.
I would suggest investigating further with a single application with a
smaller input set.
2011/5/8 Muhammad abid Mughal <mabidm_pieas at yahoo.com>:
> hi guys,
> Hope you all doing good.I am working on data prefetching of
> PARSEC applications.I calculated the opportunity i.e. Max. number of misses
> that can be removed from base system , for PARSEC on my base system.Then i
> ran some experiments using enhanced system i.e. system having prefetcher
> and found that number of misses removed were greater than Opportunity which
> is not possible.This is all happening because of indeterminism in PARSEC
> applications.I see that base system and enhanced one have different
> instruction counts.Running the same simulation multiple times(and getting
> average) with enhanced system is not feasible .Even then the average of
> misses(removed in enhanced system) across multiple simulations of enhanced
> would not be less than Opportunity calculated using Base system(This is my
> guess because Opportunity is 65% whereas enhanced system tells 84%.)
> Hope to see response from someone.
> Muhammad abid
> parsec-users mailing list
> parsec-users at lists.cs.princeton.edu
parsec-users mailing list
parsec-users at lists.cs.princeton.edu
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the parsec-users