[parsec-users] speed-up of parsec
cbienia at CS.Princeton.EDU
Thu Jul 31 12:42:15 EDT 2008
(1) You should measure speedup only for the ROI (Region of Interest). If
the main thread of canneal has plenty of instructions then you have
accidentally included the initialization phase as well - the main thread for
that program should be completely idle during the ROI. You can use PARSEC
Hooks to instrument the beginning and the end of the ROI.
(2) On real machines you might get worse speedups than those in Figure 1.
Keep in mind that we report upper speedup bounds in the report. In practice,
limitations such as memory bandwidth can give you worse speedups.
From: parsec-users-bounces at lists.cs.princeton.edu
[mailto:parsec-users-bounces at lists.cs.princeton.edu] On Behalf Of Qingyuan
Sent: Thursday, July 31, 2008 5:14 AM
To: PARSEC Users
Subject: [parsec-users] speed-up of parsec
In the parsec report paper, you caculated the achievable speed-up of parsec
benchmarks in figure 1. May I know how did you caculate the speed-ups of
Bodytrack and Canneal, since both of them have plenty of instructions in the
main thread which cannot be ignored, and it seems that in Bodytrack, the
main thread is running concurently while the children threads are woking. I
cann't get your data especially on Canneal...
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the parsec-users