home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.benchmarks
- Path: sparky!uunet!zaphod.mps.ohio-state.edu!wupost!udel!rochester!cantaloupe.srv.cs.cmu.edu!netnews-2.srv.cs.cmu.edu!moss
- From: moss@cs.cmu.edu (Eliot Moss)
- Subject: Re: DEC ALPHA Performance Claims
- In-Reply-To: silverm@bcstec.ca.boeing.com's message of 19 Nov 92 05:51:40 GMT
- Message-ID: <MOSS.92Nov20102522@CRAFTY.cs.cmu.edu>
- Followup-To: comp.benchmarks
- Sender: news@cs.cmu.edu (Usenet News System)
- Nntp-Posting-Host: crafty.fox.cs.cmu.edu
- Reply-To: moss@cs.cmu.edu
- Organization: Dept of Comp and Info Sci, Univ of Mass (Amherst)
- References: <BxH7s7.5Cv@inews.Intel.COM> <4248@bcstec.ca.boeing.com>
- <1992Nov16.174912.22905@ryn.mro4.dec.com> <4288@bcstec.ca.boeing.com>
- Date: Fri, 20 Nov 1992 15:25:22 GMT
- Lines: 50
-
- >>>>> On 19 Nov 92 05:51:40 GMT, silverm@bcstec.ca.boeing.com (Jeff Silverman)
- >>>>> said:
-
- [ some other stuff omitted ... ]
-
- Jeff> I have also received several message from several people at customer
- Jeff> sites. A lot of them are number crunchers, and they work in FORTRAN.
- Jeff> Why FORTRAN? Because it's fast (FORTRAN tends to be faster than C
- Jeff> because you don't have to build argument lists on the stack - whether
- Jeff> this is important or not, I'm not sure - I like to program in "C"). So
- Jeff> they aren't so concerned with the speed of the chip, they're concerned
- Jeff> with the speed of the computer, which is the compiled code, memory,
- Jeff> disk, the whole ball of wax.
-
- That may be true, but one should pass most arguments in registers in any case
- (unless you have long parameter lists), and if they have to be dumped to
- memory, putting them in the stack is no more costly than putting them in
- static locations. Perhaps the FORTRANers are objecting to the stack pointer
- manipulating instructions? Hardly seems a big cost to me. More likely that
- FORTRAN optimizers are ahead on really agressive loop optimizations and such.
- But this is not the main point I wanted to address ....
-
- Jeff> I also got one message which was fascinating - it was pointed out to me
- Jeff> that the activities associated with having more demand for RAM than RAM
- Jeff> are a major performance drag. Paging is bad news if you are into speed,
- Jeff> swapping is even worse. So how about a computer system where there is
- Jeff> enough physical RAM so that everything that must be RAM resident is, and
- Jeff> it doesn't swap or page. This operating system would be something like
- Jeff> RT-11, but it would be called RT-64 or similar. I remember that on the
- Jeff> DECsystem-10 (KA-10 CPU) you could get a performance goose by turning
- Jeff> off the relocation logic, but that made it tough to run TOPS-10.
-
- If you're sufficiently privileged, you can nail down your pages. Modern
- machines' relocation logic has been worked on to make it fast, though I can't
- say if the 21064 manages to run as fast as it would without relocation logic.
- I suspect paging/swapping can be taken care of. What's going to be harder is
- getting good cache performance. One needs a compiler that knows how to
- reorganize loops to get more cache hits, etc. Some techniques have been
- developed that do help a lot of loops, but I don't know the extent to which
- they've made it into commerically available compilers.
-
- --
-
- J. Eliot B. Moss, Associate Professor Visiting Associate Professor
- Department of Computer Science School of Computer Science
- Lederle Graduate Research Center Carnegie Mellon University
- University of Massachusetts 5000 Forbes Avenue
- Amherst, MA 01003 Pittsburgh, PA 15213-3891
- (413) 545-4206, 545-1249 (fax) (412) 268-6767, 681-5739 (fax)
- Moss@cs.umass.edu Moss@cs.cmu.edu
-