Lahey Support
08-15-2003, 01:33 AM
Hi Klaus:
I've just gone through a similar exercise with a very large Monte Carlo
code.
I compiled the code on both a Pentium system (ancient P-1 400 MHz) and 1.2
Ghz AMD (K-7) and ran a very thorough test suite provided by the authors.
Perfect agreement (to the last digit) with approved answers on both
machines. I also have trial use of a 800 MHz P-3, on which I shall now try
the same test, as soon as I can, thanks to your question.
Many years ago, I once (and only ever once) saw disagreement between
numerical results from the same compiled code run on different machines - it
turned out to be a hardware bug on the motherboard of one of them, and could
be isolated by running the code twice in succession after a cold boot. Got
different answers. I actually managed to find the offending statement in
the code, and traced it to an incorrect hardware implementation of the
compiler's (LF77) treatment of stack overflow. I have also recently had
problems that appear/disappear on the same machine depending on whether that
machine has been freshly booted, although these are more serious than small
arithmetic differences.
On the off-chance you are seeing a similar effect, try booting and compare
the answers before and after. Also, try compiling the code on both machines
separately and see what happens.
Good luck - this could turn out to be a very useful investigation. In fact
if it is possible to wrap up the executable and the offending input file
into a simple package that I could try on some of the test machines I have
here, I'd be very interested, and would try and get trial access to a P-4
for a test.
Otherwise, could you please keep us informed - I'd like to know what you
find, since I am on the verge of buying my first AMD machine, but with some
trepidation.
Mike Milgram
Klaus Bigalke wrote:
> Hallo,
>
> has somebody experience in comparing the results of numerical procedures
> run on AMD Athlon / INTEL P4 systems?
>
> Working with Athlon for years I've just got a new P4 system and run
> the same numerical code (mesoscale meteorological simulation model,
> compiled with Lf95 5.5j)
> a) on an AMD Athlon 850MHz, 512 Mb RAM
> b) on an Intel P4 1800MHz, 512 Mb RAM.
> The results are only "nearly" identical. The differences are of an order
> I would expect when running the code eg. on a Cray system (where
> not only the system would be different but also the compiler, numerical
> precision etc.).
> I've never seen any differences between PC's - regardless of the
> processor (AMD or INTEL).
>
> Klaus
>
> METCON Umweltmeteorologische Beratung Dr. Klaus Bigalke
> Jappopweg 9h, D-25421 Pinneberg
> Tel. 04101/6938-56 Fax: 04101/6938-57 Mobil: 0171/6249543
> Email: [address removed] Web: www.metcon-umb.de
----------------------------------------------------------
To unsubscribe, send to [address removed] the following
as the first and only line of the message body:
unsubscribe fortran
----------------------------------------------------------
I've just gone through a similar exercise with a very large Monte Carlo
code.
I compiled the code on both a Pentium system (ancient P-1 400 MHz) and 1.2
Ghz AMD (K-7) and ran a very thorough test suite provided by the authors.
Perfect agreement (to the last digit) with approved answers on both
machines. I also have trial use of a 800 MHz P-3, on which I shall now try
the same test, as soon as I can, thanks to your question.
Many years ago, I once (and only ever once) saw disagreement between
numerical results from the same compiled code run on different machines - it
turned out to be a hardware bug on the motherboard of one of them, and could
be isolated by running the code twice in succession after a cold boot. Got
different answers. I actually managed to find the offending statement in
the code, and traced it to an incorrect hardware implementation of the
compiler's (LF77) treatment of stack overflow. I have also recently had
problems that appear/disappear on the same machine depending on whether that
machine has been freshly booted, although these are more serious than small
arithmetic differences.
On the off-chance you are seeing a similar effect, try booting and compare
the answers before and after. Also, try compiling the code on both machines
separately and see what happens.
Good luck - this could turn out to be a very useful investigation. In fact
if it is possible to wrap up the executable and the offending input file
into a simple package that I could try on some of the test machines I have
here, I'd be very interested, and would try and get trial access to a P-4
for a test.
Otherwise, could you please keep us informed - I'd like to know what you
find, since I am on the verge of buying my first AMD machine, but with some
trepidation.
Mike Milgram
Klaus Bigalke wrote:
> Hallo,
>
> has somebody experience in comparing the results of numerical procedures
> run on AMD Athlon / INTEL P4 systems?
>
> Working with Athlon for years I've just got a new P4 system and run
> the same numerical code (mesoscale meteorological simulation model,
> compiled with Lf95 5.5j)
> a) on an AMD Athlon 850MHz, 512 Mb RAM
> b) on an Intel P4 1800MHz, 512 Mb RAM.
> The results are only "nearly" identical. The differences are of an order
> I would expect when running the code eg. on a Cray system (where
> not only the system would be different but also the compiler, numerical
> precision etc.).
> I've never seen any differences between PC's - regardless of the
> processor (AMD or INTEL).
>
> Klaus
>
> METCON Umweltmeteorologische Beratung Dr. Klaus Bigalke
> Jappopweg 9h, D-25421 Pinneberg
> Tel. 04101/6938-56 Fax: 04101/6938-57 Mobil: 0171/6249543
> Email: [address removed] Web: www.metcon-umb.de
----------------------------------------------------------
To unsubscribe, send to [address removed] the following
as the first and only line of the message body:
unsubscribe fortran
----------------------------------------------------------