From owner-freebsd-alpha Tue Sep 19 8:43:53 2000 Delivered-To: freebsd-alpha@freebsd.org Received: from blighty.com (vix.samspade.org [204.152.186.148]) by hub.freebsd.org (Postfix) with ESMTP id 8544337B423 for ; Tue, 19 Sep 2000 08:43:51 -0700 (PDT) Received: by blighty.com (Postfix, from userid 1000) id 4C724533; Tue, 19 Sep 2000 08:43:59 -0700 (PDT) Date: Tue, 19 Sep 2000 08:43:59 -0700 From: Steve Atkins To: freebsd-alpha@freebsd.org Subject: Re: alpha ev56 contra ev67 Message-ID: <20000919084359.B12931@blighty.com> References: <5ED89301AE60D311AAD500508B0EF55F50E070@dd-mail.intern.dd.dk> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii X-Mailer: Mutt 1.0.1i In-Reply-To: <5ED89301AE60D311AAD500508B0EF55F50E070@dd-mail.intern.dd.dk>; from bbl@dk.damgaard.com on Tue, Sep 19, 2000 at 03:54:26PM +0200 Sender: owner-freebsd-alpha@FreeBSD.ORG Precedence: bulk X-Loop: FreeBSD.org On Tue, Sep 19, 2000 at 03:54:26PM +0200, Bjarne Blichfeldt wrote: > Gentlemen, > > Not directly FreeBSD, but close. > > What are the changes from ev56 to ev67? Out of order execution is the biggie, but the cache increased in size too - from 8/8+96 to 64/64 > Specifically, will I see any performance increase from compiling a > program on a ev67, compared to compiling on a ev56 and running the > same program on a ev67. I assume you mean optimise for EV6/EV5 rather than compile on EV6/EV5? > ex. > 1. compile program AA on a ev56, execute on a ev56 = performance XX > 2. compile program AA on a ev56, execute on a ev67 = performance XX + some > % > 3. compile program AA on a ev67, execute on a ev67 = performance ?? > compared to 2. > > Depends on the program, I know, but in general non obligated terms ;-) > Compiling the program is done with gcc 2.95.2 in this scenario. Generally, a program that is compiler-optimised for EV5 will run nearly as well on an EV6 as a program compiler-optimised for EV6. Code optimised for an in-order machine will run nicely on an OOO machine. There'll always be some improvement optimising for EV6, and occasionally the improvement is large. Performance differences will depend on lots of things. If it is memory intensive and the inner loop fits into a 64k cache but not an 8k cache you'll see a big speedup. If your application smells like SPECint or SPECfp you may see a c. 2x performance improvement. > Does anybody, by the way, have any idea of how the gcc compiler fares, > compared > to Compaq's own compiler on the alpha platform ? Poorly. Compaqs compiler/linker/optimiser suite is very good. Gcc isn't that bad, though. The effect is less spectacular on an EV6, as the OOO makes clever scheduling less valuable. Cheers, Steve To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-alpha" in the body of the message