[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-smp
Subject:    PIII versus Athlon
From:       Herbert Huber <Herbert.Huber () lrz-muenchen ! de>
Date:       1999-10-27 8:30:11
[Download RAW message or body]


> I'd run stream_d earlier using a binary of forgotten origin.  I've
re-built
> and rerun.

> All of the results below are on an Intel N440BX with two 500 MHz
PIII's, 128
> MB of PC100 SDRAM, compiled with -O2, running 2.2.12 Linux, with
stream_d
> built to use 91.6 MB of memory (Array size = 4000000).  The system was

> otherwise idle.


> egcs-2.91.66, cpu timer
> -----------------------
> Function      Rate (MB/s)   RMS time     Min time     Max time
> Copy:         320.0000       0.2000       0.2000       0.2000
> Scale:        320.0000       0.2041       0.2000       0.2100
> Add:          384.0000       0.2581       0.2500       0.2700
> Triad:        300.0000       0.3310       0.3200       0.3400

> egcs-2.91.66, wall clock timer
> ------------------------------
> Function      Rate (MB/s)   RMS time     Min time     Max time
> Copy:         320.0208       0.2004       0.2000       0.2014
> Scale:        318.4571       0.2013       0.2010       0.2025
> Add:          372.0700       0.2594       0.2580       0.2690
> Triad:        294.3115       0.3292       0.3262       0.3459

To clear the situation a little bit. My first PIII-results have been
measured on a dual system (Asus P2B-DS, 512 MB RAM), running one big
gaussian job:

> Function      Rate (MB/s)   RMS time     Min time     Max time
> Copy:         253.5778       0.0636       0.0631       0.0647
> Scale:        248.1850       0.0649       0.0645       0.0653
> Add:          297.6852       0.0816       0.0806       0.0826
> Triad:        296.7909       0.0820       0.0809       0.0843

So the total memory bandwith was shared between two processes!

The results below are measured on an Asus P3B-F with one 500 MHz PIII,
128 MB PC100 SDRAM, PGI 3.1 Fortran compiler, compiled with -fast -tp
p6, running linux-2.2.12. stream_d was built to use 91 MB of memory
(n=4000000). In this case the system was otherwise idle.

cpu timer
----------------------------------------------------
Function     Rate (MB/s)  RMS time   Min time  Max time
Copy:        320.0000      0.2020      0.2000      0.2100
Scale:       320.0000      0.2000      0.2000      0.2000
Add:         384.0000      0.2550      0.2500      0.2600
Triad:       384.0000      0.2520      0.2500      0.2600

wall timer
 ----------------------------------------------------
Function     Rate (MB/s)  RMS time   Min time  Max time
Copy:        318.2211      0.2013      0.2011      0.2022
Scale:       318.3176      0.2011      0.2011      0.2012
Add:         377.6108      0.2544      0.2542      0.2549
Triad:       380.6819      0.2524      0.2522      0.2527

You can download the ziped tar file containing the binaries and their
md5 via anonymous ftp from:
ftp.lrz-muenchen.de:/transfer/streamdir/stream.tar.gz

The md5 sum of this file is: e64531ed59666d76c3b99ec6d2973e6f

The binaries are built on an glibc2.0 based system. They do not run on
systems based on glibc2.1.

/Herbert




-
Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/dmentre/smp-howto/
To Unsubscribe: send "unsubscribe linux-smp" to majordomo@vger.rutgers.edu

[prev in list] [next in list] [prev in thread] [next in thread]