On Thu, Jul 24, 2003 at 11:50:00AM +0200, Matthias Urlichs wrote: > Hi, Andrew Suffield wrote: > > On Thu, Jul 24, 2003 at 10:18:22AM +0200, Matthias Urlichs wrote: > >> Well, when the glibc people had this discussion (the switch to ELF), > >> the performance penalty was found to be on the order of 5%. > >> > ... by testing with somewhat typical programs. > > >> I don't know whether modern CPUs' register aliasing hardware changes > >> that number. > > > > On anything more recent than a 486, the answer is not a percentage, > > It is -- if you do the same tests, and compare. Well, you can get an *answer* that way, it's just not very accurate or useful :P > > it's a research paper. Just to _describe_ the performance implications. > > ... assuming you want to translate the results to a different set of > tests. > > ... or, for that matter, if you want to interpret the variance you'll get > in the test results, assuming it's significant, which it probably will be. You generally get a significant variance when you modify the input data. Although that doesn't always hold. -- .''`. ** Debian GNU/Linux ** | Andrew Suffield : :' : http://www.debian.org/ | `. `' | `- -><- |
Attachment:
pgphuFG08Nj5h.pgp
Description: PGP signature