Re: Floating point performance of UNIX workstations.
From: Lord Isildur (isildur_at_andrew.cmu.edu)
Date: 06/09/03
- Next message: Chris Morgan: "Re: Floating point performance of UNIX workstations."
- Previous message: GccTx: "System Recover Help"
- In reply to: Chip Coldwell: "Re: Floating point performance of UNIX workstations."
- Next in thread: Adam Price: "Re: Floating point performance of UNIX workstations."
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ] [ attachment ]
Date: Mon, 9 Jun 2003 14:09:01 -0400 (EDT)
I just tested it on a 433au running netbsd 1.4.1, and got 863 sec, without
tweaking any compiler flags or anything. the system has 2 megs of cache, 640
megs core, and some random 10k rpm disks. gcc version is 2.91.60. one thing i
noticed is that the makefiles generated did not turn on ANY optimization,
not even -O! i'm recompiling with -O3 now. that just finished, in 385 sec.
At the same time, i decided to compile and test it on a dec3000/600 running
netbsd 1.5.2. that is, of course, an old-school alpha, 21064 at 175 MHz.
with no optimization, it ran in 2764 sec. With -O3, it ran in 1276 sec.
Looking at it again, i notice that the configure script doesnt seem to
be able to figure out what kind of machine it is, and apparently is defaulting
to a 'safe' configuration or sorts.. including no optimization.
Isildur
On Mon, 9 Jun 2003, Chip Coldwell wrote:
> In comp.sys.dec Dr. David Kirkby <drkirkby@ntlworld.com> wrote:
>
> > What I can't understand is quite why the Dec Alpha seems to
> > fair so very badly, when I run a floating point intensive simulation
> > on these systems.
>
> Which compiler were you using? Ages ago, I did benchmarks comparing
> gcc generated code with the code generated by the Digital compilers.
> The difference was huge; 3x or 4x speedup was not unusual.
>
> Also, be careful with floating point on the Alpha. It supports two
> floating point formats, IEEE 754 (standard) and VAX (for backward
> combatibility) as well as software emualtion:
>
> -mno-soft-float -mfp-reg -mieee
>
> might be a good set of compiler flags to try.
>
> Chip
>
> --
> Charles M. "Chip" Coldwell
> "Turn on, log in, tune out"
>
>
- Next message: Chris Morgan: "Re: Floating point performance of UNIX workstations."
- Previous message: GccTx: "System Recover Help"
- In reply to: Chip Coldwell: "Re: Floating point performance of UNIX workstations."
- Next in thread: Adam Price: "Re: Floating point performance of UNIX workstations."
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ] [ attachment ]
Relevant Pages
|