[parsec-users] branch intensive benchmark

Fedorova, Julia Julia.Fedorova at intel.com
Fri Jul 16 03:06:51 EDT 2010

Hi Mahmood

I won't be so certain on this.

Recently I was able to speed up fluidanimate by ~ 20% improving code around 2 mispredicted branches.
(I've sent that patch to PARSEC list  - pls let me know if you want it)

The fix was not improving the quality of the prediction
but rather avoiding the execution of the poorly predicted branches ,when it is possible, and a number of instructions around it.

As a result - I got significant absolute decrease in number of mispredicted branches and some decrease in number of overall retired instructions. And as final result ~ 20% improvement of elapsed time.

It looks like Figure 8 estimates Branch misprediction impact as 10% (for fluidanimate).
But my estimations on Intel Core i7 - showed ~ 30%.
And the fix I made kind of closer to my number. (while more analysis should be done on this)

So before the fix and even after - the branch misprediction was and is still glaring problem at fluidanimate.

Hope this was useful.

Regards, Julia
From: parsec-users-bounces at lists.cs.princeton.edu [mailto:parsec-users-bounces at lists.cs.princeton.edu] On Behalf Of Mahmood Naderan
Sent: Friday, July 16, 2010 10:32 AM
To: PARSEC Users
Subject: Re: [parsec-users] branch intensive benchmark

The only thing related to branch mispredict is Figure 8. From that I understand branch mis predict isn't a major bottleneck. Thanks,

// Naderan *Mahmood;

From: Christian Bienia <cbienia at CS.Princeton.EDU>
To: PARSEC Users <parsec-users at lists.cs.princeton.edu>
Sent: Fri, July 16, 2010 12:54:24 AM
Subject: Re: [parsec-users] branch intensive benchmark
Hi Naderan,

I haven't characterized the branch behavior of PARSEC myself, but Major Bhadauria did some work on single-thread characteristics. You might find some relevant information in his publication "Understanding PARSEC Performance on Contemporary CMPs["


From: parsec-users-bounces at lists.cs.princeton.edu [mailto:parsec-users-bounces at lists.cs.princeton.edu] On Behalf Of Mahmood Naderan
Sent: Thursday, July 15, 2010 1:31 PM
Subject: [parsec-users] branch intensive benchmark

Which benchmark is "branch instruction" intensive?

// Naderan *Mahmood;

Closed Joint Stock Company Intel A/O
Registered legal address: Krylatsky Hills Business Park, 
17 Krylatskaya Str., Bldg 4, Moscow 121614, 
Russian Federation

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.cs.princeton.edu/pipermail/parsec-users/attachments/20100716/95ebf61e/attachment-0001.html>

More information about the parsec-users mailing list