home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.unix.internals
- Path: sparky!uunet!news.univie.ac.at!doris!alex
- From: alex@doris.par.univie.ac.at (Alexander Celic)
- Subject: Big problem with performance measuring on a INTEL HYPERCUBE
- Message-ID: <1993Jan2.185750.15342@newssrv.edvz.univie.ac.at>
- Sender: alex@doris (Alexander Celic)
- Nntp-Posting-Host: doris.par.univie.ac.at
- Organization: University of Vienna
- Date: Sat, 2 Jan 1993 18:57:50 GMT
- Lines: 12
-
-
- Hi, there!
-
- I've been measuring performance of some Fortran 77 commands on an
- Intel's HYPERCUBE. The data I've become is very unstable (read: one
- kernel takes ALLWAYS same other time to execute; the difference between
- two different meassurements can be up to a factor 4 !!!)
-
- Of course, the main source of this problem lies on pipeline and
- cache, but is there allso something more? If somebody had the same problem,
- I'd really like to know how he (she?) solved it. I couldn't think of anything
- but taking the middle result. Has anybody found some other way?
-