How accurate are the branch-prediction numbers?

Mayan Moudgill <moudgill@cs.cornell.EDU>
Thu, 23 Sep 1993 17:40:55 GMT

          From comp.compilers
Related articles
How accurate are the branch-prediction numbers? moudgill@cs.cornell.EDU (Mayan Moudgill) (1993-09-23)
| List of all articles for this month |
Newsgroups: comp.compilers
From: Mayan Moudgill <moudgill@cs.cornell.EDU>
Keywords: optimize, question
Organization: Cornell Univ. CS Dept, Ithaca NY 14853
Date: Thu, 23 Sep 1993 17:40:55 GMT

Most references on branch prediction claim to get less than 10%
mispredicted branches (using both dynamic and static techniques) for their
programs. However, in talking with people who measure system performance
directly, the mispredicted number seems to be closer to 30-35%. This, of
course, includes I/O and interrupt handlers, O/S code and a code mix that
includes a fair amount of data-processing code.


Any ideas who's telling the truth? Is it that the benchmarks used don't
accurately measure O/S code? Or is it an artifact of the machine+code mix
being measured?


:)
Mayan
--

Post a followup to this message

Return to the comp.compilers page.
Search the comp.compilers archives again.