Home > Archive > Fortran > May 2005 > best cpu
You are viewing an archived Text-only version of the thread.
To view this thread in it's original format and/or if you want to reply to
this thread please [click here]
|
|
| Tony199 2005-05-14, 4:05 pm |
| hello.
i wrote over 5000 lines of code, and compilation needs about 30seconds,
the programm code needs about 320seconds to execute. i am working on
amd 2500+ and code is optimized for amd athlon. but i need the code to
work faster. i would like to change the cpu / complete pc system. what
cpu is best/fastest for executing fortran code? i think the FPU must be
good. Xeon? Pentium 4 EE? Opteron? Athlon FX? Itanium? any ideas?
thanks!
| |
| Tony199 2005-05-14, 4:05 pm |
| i have to add the following:
same code on a single intel xeon 2200Mhz needs about 352seconds. ram is
on both systems over 1GB. would be very nice to recieve some ideas
about making computing faster.
regards
| |
| Tim Prince 2005-05-14, 4:05 pm |
| Tony199 wrote:
> hello.
> i wrote over 5000 lines of code, and compilation needs about 30seconds,
> the programm code needs about 320seconds to execute. i am working on
> amd 2500+ and code is optimized for amd athlon. but i need the code to
> work faster. i would like to change the cpu / complete pc system. what
> cpu is best/fastest for executing fortran code? i think the FPU must be
> good. Xeon? Pentium 4 EE? Opteron? Athlon FX? Itanium? any ideas?
> thanks!
The lack of information here makes this look like a troll.
If you don't parallelize, any of those which support MP would be off target.
If you do, the same program may perform best on any of those models which
support MP, depending on the data.
--
Tim Prince
| |
| Steve Kargl 2005-05-14, 4:05 pm |
| Tony199 wrote:
> hello.
> i wrote over 5000 lines of code, and compilation needs about 30seconds,
> the programm code needs about 320seconds to execute. i am working on
> amd 2500+ and code is optimized for amd athlon. but i need the code to
> work faster. i would like to change the cpu / complete pc system. what
> cpu is best/fastest for executing fortran code? i think the FPU must be
> good. Xeon? Pentium 4 EE? Opteron? Athlon FX? Itanium? any ideas?
> thanks!
>
Have you used a profiler to determine where the code spends it time?
The choice of a bad algorithm can be a killer.
--
steve
| |
| Tony199 2005-05-14, 4:05 pm |
| sorry i am not a troll. where is the problem? lack of information? what
do you want more? well the code, yes is very huge, but the code is also
very sientific, the cpu has to compute a lot of doubles.
with my current knowledge, i think the prozess (theard) is only one, so
MP would not help me. but i have seen that the same code runns very
different if the cpu is intel or amd. i thought my code ( coded with
intel fortran compiler 8.1) would run on an intel xeon a lot faster
than on an amd athlon 2500+. but thats not. so i ask now here if you
know where fortran code is computed faster than on my current system.
yes sure, itanium was a bit stupid. it is nearly not affordable.
i looked around and checked some cpus with sisoft sandra. and the
benchmark system gave me some information about FPU, ALU, MIPS, GFLOPS
etc. maybe you know which value is important to speed up code. i
thought, the faster the FPU is the faster the fortran code weill
speeded up.
and i am not a troll :-)
Tim Prince schrieb:
> Tony199 wrote:
>
30seconds,[color=darkred]
on[color=darkred]
to[color=darkred]
what[color=darkred]
must be[color=darkred]
> The lack of information here makes this look like a troll.
> If you don't parallelize, any of those which support MP would be off
target.
> If you do, the same program may perform best on any of those models
which
> support MP, depending on the data.
> --
> Tim Prince
| |
| Tony199 2005-05-14, 4:05 pm |
| sorry, no i have not used a profiler. i just upgrade from Compaq 6.6 to
intel fortran 8.1 and everything seems to be a bit different. where can
i find that profiler? sorry. :-)
| |
| Gordon Sande 2005-05-14, 4:05 pm |
|
Tony199 wrote:
> sorry i am not a troll. where is the problem? lack of information? what
> do you want more? well the code, yes is very huge, but the code is also
> very sientific, the cpu has to compute a lot of doubles.
> with my current knowledge, i think the prozess (theard) is only one, so
> MP would not help me. but i have seen that the same code runns very
> different if the cpu is intel or amd. i thought my code ( coded with
> intel fortran compiler 8.1) would run on an intel xeon a lot faster
> than on an amd athlon 2500+. but thats not. so i ask now here if you
> know where fortran code is computed faster than on my current system.
> yes sure, itanium was a bit stupid. it is nearly not affordable.
>
> i looked around and checked some cpus with sisoft sandra. and the
> benchmark system gave me some information about FPU, ALU, MIPS, GFLOPS
> etc. maybe you know which value is important to speed up code. i
> thought, the faster the FPU is the faster the fortran code weill
> speeded up.
>
> and i am not a troll :-)
>
The most constructive answer to how to lower your execution time is
to improve your algorithm. The most polished n^2 sort (bubble, merge
whatever) will be left in the dust by even a minimally competent n
log(n) sort (quick, tree, whatever). Sometimes algorithm choice means
not getting into trouble with paging or caching issues.
You have done a superb imitation of a clown who thinks that a highly
polished lousy algorithm ever stands a chance of being efficient.
The requests for more information are trying to find out what type
of problem you are dealing with so that efficient algorithms can
be suggested. Your response reinforces the notion that you are
almost surely totally ignorant of the many variants on algorithm
analysis which have made many problems tractable. Unfortunately
you have also chosen to associate yourself with trolls and the
like by choosing to have only a handle rather than a a full name.
Until you have a realistic problem statement beyond 5000 lines of
code that use double precision you will get no useful advice.
By the way, 5000 lines of code is not very big for this newsgroup.
>
> Tim Prince schrieb:
>
>
> 30seconds,
>
>
> on
>
>
> to
>
>
> what
>
>
> must be
>
>
> target.
>
>
> which
>
>
>
| |
| Tony199 2005-05-14, 4:05 pm |
| In any case, thanks for any suggestions, Mr Gordon Sande. Next time
I'll repsect the netiquette, I am sorry. The code is part of a diss and
I really know about algorisms but I wanted to know something about CPUs
and not coding. So thanks, to comment my beheaviour in 17 lines... (wt)
| |
| Tim Prince 2005-05-14, 4:05 pm |
| Tony199 wrote:
> sorry, no i have not used a profiler. i just upgrade from Compaq 6.6 to
> intel fortran 8.1 and everything seems to be a bit different. where can
> i find that profiler? sorry. :-)
ifort 8.1 on linux should work well with gprof. Does your statement imply
you are limited to Windows? Try VTune.
--
Tim Prince
| |
| Tony199 2005-05-14, 4:05 pm |
| thank you. yes i am limited to windows only. vtune, i'll try it. thanks
very much.
btw: i have found that parallelizer and openMP. so far i understand
that, it means that the parallelizer makes code that can be handled by
muli-threading. i think that will boost up the code. i saw that you
have to use especial compilation codes to especialize cpus. I will try
it out. thank you so far!!
| |
| Steven G. Kargl 2005-05-14, 7:13 pm |
| In article <1116091416.022483.194580@g47g2000cwa.googlegroups.com>,
"Tony199" <spam1@servido.net> writes:
> In any case, thanks for any suggestions, Mr Gordon Sande. Next time
> I'll repsect the netiquette, I am sorry. The code is part of a diss and
> I really know about algorisms but I wanted to know something about CPUs
> and not coding. So thanks, to comment my beheaviour in 17 lines... (wt)
Gordon is right. We need much more information than
you have given. 5000 lines of code and double precision
really don't mean much. What you have told is equivalent
to one of us calling our auto mechanic and saying "my car
won't run and it has premium gasoline."
Are you doing numerical integration, large matrix problems,
finite difference, sorting and hash computations? Are you
using nested loops with row-major indexing in the inner
loop? Are you doing a massive amount of IO?
--
Steve
http://troutmask.apl.washington.edu/~kargl/
| |
| Tim Prince 2005-05-14, 7:13 pm |
| Tony199 wrote:
> thank you. yes i am limited to windows only. vtune, i'll try it. thanks
> very much.
> btw: i have found that parallelizer and openMP. so far i understand
> that, it means that the parallelizer makes code that can be handled by
> muli-threading. i think that will boost up the code. i saw that you
> have to use especial compilation codes to especialize cpus. I will try
> it out. thank you so far!!
Sorry, forgot to mention that VTune will support only the Intel CPU, while
gprof (on linux, or with gfortran on cygwin) will allow you to compare
performance with that other brand.
Yes, auto parallelizer will perform only a fraction of what can be done with
OpenMP directives. OpenMP should make a big difference on dual core or
dual CPU platforms, if your source code is suitable.
--
Tim Prince
| |
| glen herrmannsfeldt 2005-05-15, 3:57 am |
| Tony199 wrote:
> sorry i am not a troll. where is the problem? lack of information? what
> do you want more? well the code, yes is very huge, but the code is also
> very sientific, the cpu has to compute a lot of doubles.
Yes, then you want a fast floating point unit...
(snip)
> and i am not a troll :-)
We believe you, but it was hard to tell from the original post.
You didn't even mention that it was doing floating point.
My favorite are the SPEC programs. While you might start looking
just at the SPEC-FP numbers, you could also try to decide which of
the SPEC benchmark programs is most similar to what you are trying
to do. Many of them are Fortran.
Another important difference is between x87 instructions and the
newer SSE2 instructions. Some compilers use one set and some the
other, and the timing may be very different.
-- glen
| |
| beliavsky@aol.com 2005-05-15, 3:57 am |
| Tony199 wrote:
> sorry, no i have not used a profiler. i just upgrade from Compaq 6.6
to
> intel fortran 8.1 and everything seems to be a bit different. where
can
> i find that profiler? sorry. :-)
If you cannot find the information in the documentation of the Intel
Fortran compiler for Windows, the best place to ask questions specific
to this compiler is the forum for the "Intel Fortran Compiler for
Windows & Visual Fortran" at
http://softwareforums.intel.com/ids/board?board.id=5 .
Are you aware of intrinsic functions such as CPU_TIME, which can be
used to measure the time taken by various parts of a program?
P.S. Welcome to comp.lang.fortran. Most people here try to help, even
when scolding :).
| |
| Michael Gray 2005-05-16, 8:57 am |
| On 14 May 2005 09:42:03 -0700, "Tony199" <spam1@servido.net> wrote:
>sorry, no i have not used a profiler. i just upgrade from Compaq 6.6 to
>intel fortran 8.1 and everything seems to be a bit different. where can
>i find that profiler? sorry. :-)
http://www.intel.com/software/produ...e/vpa/index.htm
|
|
|
|
|