Parallel computing on dual-processor pentium 
Author Message
 Parallel computing on dual-processor pentium

Dear all,

Could anyone please tell me if it is possible to run parallel
applications on dual-processor pentiums, running under Linux,
specifically applications written in HPF90 (although I'd be interested
to hear about MPI as well).

If it is possible what additional software would I require (eg message
passing libraries, compilers...) and could anyone reccomend specific
compilers.

Many thanks,

Stuart Caunt.



Wed, 18 Jun 1902 08:00:00 GMT  
 Parallel computing on dual-processor pentium
Quote:

> Dear all,

> Could anyone please tell me if it is possible to run parallel
> applications on dual-processor pentiums, running under Linux,
> specifically applications written in HPF90 (although I'd be interested
> to hear about MPI as well).

Well, yes!

Although I don't know about ANY HPF compiler that does "real"
parallelization on SMP machines. All I've herad of is a parallelization
mechanism with underlying MPI function calls. So you'll need an MPI
implementation besides the HPF compiler.

MPICH is the most common freely available MPI implementation, but it has
some problems on some versions of Linux and some fortran compilers
regarding using SMP techniques (say: shared memory).

Quote:
> If it is possible what additional software would I require (eg message
> passing libraries, compilers...) and could anyone reccomend specific
> compilers.

I have very good performance with NAGware's f95 (available for
Linux-2.2.x/glibc-2.1) and MPICH or LAM/MPI. Have a look at the
Parallel-Processing HOWTO's available at a mirror of your choice.

Or ask a search engine of your choice for MPI, HPF90 and Linux :-)
(Sorry, but this IS a FAQ!)

cul8r,
         Frank
--
  while (!asleep()) sheep++;



Wed, 18 Jun 1902 08:00:00 GMT  
 Parallel computing on dual-processor pentium

Quote:

> Dear all,

> Could anyone please tell me if it is possible to run parallel
> applications on dual-processor pentiums, running under Linux,
> specifically applications written in HPF90 (although I'd be interested
> to hear about MPI as well).

> If it is possible what additional software would I require (eg message
> passing libraries, compilers...) and could anyone reccomend specific
> compilers.

> Many thanks,

> Stuart Caunt.

If all you're after is using a two way SMP, then OpenMP
(http://www.openmp.org) may be more appropriate than
doing a full distributed memory message pasing re-implementation of
your code.

I believe that the PGI compiler (http://www.pgroup.com) supports OpenMP
code on Linux (though I haven't tried it, and there may be other compilers
which also do).

-- Jim


Etnus, Inc.     +44 117 9071438
http://www.etnus.com



Wed, 18 Jun 1902 08:00:00 GMT  
 Parallel computing on dual-processor pentium
Hi Stuart,

Quote:
> Could anyone please tell me if it is possible to run parallel
> applications on dual-processor pentiums, running under Linux,
> specifically applications written in HPF90 (although I'd be interested
> to hear about MPI as well).

> If it is possible what additional software would I require (eg message
> passing libraries, compilers...) and could anyone reccomend specific
> compilers.

My understanding is that Dakota Software (no URL springs to mind -- sorry:()
sell parallel libraries which will run under NT with the appropriate compiler
(Lahey/DVF & maybe Absoft)

Regards,

Dave.

--
 ==================================================================
 David Annetts                                       _____________
 http://www.ocs.mq.edu.au/~dannetts/                |C R C A M E T|
                                                    |-------------|
                                                    |_____        |
 CRC for Australian Mineral                         |````` \      |
 Exploration Technologies                           |`````/$\     |
 Earth & Planetary Sciences                         |````/$$$\____|
 Macquarie University, NSW 2109                     |```/$$$/.....|
 AUSTRALIA                                          |``/$$$/......|
 phone: +(1-61-2) 9850 9280, fax (1-61-2) 9850 8366  -------------
 ==================================================================



Wed, 18 Jun 1902 08:00:00 GMT  
 Parallel computing on dual-processor pentium
On Tue, 30 Nov 1999 10:29:20 +1100, David Annetts

Quote:

>My understanding is that Dakota Software (no URL springs to mind -- sorry:()
>sell parallel libraries which will run under NT with the appropriate compiler
>(Lahey/DVF & maybe Absoft)

Dakota Scientific Software was purchased by Sun and withdrew from the
NT market.  Dakota had parallelized math library routines, not a
general MP solution.

Stuart was asking about Linux - if NT was an option, then the
combination of Compaq Visual Fortran and Kuck and Associates' KAP for
OpenMP would provide this capability.  Absoft Pro Fortran MP is also
supported for use with Visual KAP for OpenMP - Lahey compilers are
not.

There doesn't seem to be much attention paid to SMP on Linux to date -
MPI and other "loosely-coupled" mechanisms seem to be most popular.
Perhaps this will change in the future.

--

Fortran Engineering
Compaq Computer Corporation, Nashua NH

Compaq Fortran web site: http://www.compaq.com/fortran



Wed, 18 Jun 1902 08:00:00 GMT  
 Parallel computing on dual-processor pentium

Quote:
> If all you're after is using a two way SMP, then OpenMP
> (http://www.openmp.org) may be more appropriate than
> doing a full distributed memory message pasing re-implementation of
> your code.

> I believe that the PGI compiler (http://www.pgroup.com) supports OpenMP
> code on Linux (though I haven't tried it, and there may be other compilers
> which also do).

I'll confirm that.

Chose the pgroup PGI compiler for just that reason.  They've got a fairly
decent auto-parallel code generation that can beat my pitiful attempts at
hand coding the parallel instructions.  And if you don't like OpenMP, you
can use their HPF compiler, too.  For a two-processor machine, I found that
HPF tended to slow things down quite a bit compared to OpenMP.

FWIW, for the code that I use (large matrix-vector multiplies), I get better
results with the PGI compiler than I do with the Lahey lf95.



Wed, 18 Jun 1902 08:00:00 GMT  
 Parallel computing on dual-processor pentium
Quote:

> Stuart was asking about Linux - if NT was an option, then the
> combination of Compaq Visual Fortran and Kuck and Associates' KAP for
> OpenMP would provide this capability.  Absoft Pro Fortran MP is also
> supported for use with Visual KAP for OpenMP - Lahey compilers are
> not.

Well, you simply don't want to run SMP Fortran programs on NT. Perhaps
this sounds like a flame to you, but there's a good article from our
local Fortran-Guru, Wilhelm Gehrke (homepage
http://www.unics.uni-hannover.de/rrzn/gehrke/, sorry, german only), in
the german "Unix" magazine iX (see http://www.heise.de/ix/inhalt.shtml,
iX 12/1999, pp.92-95)...

Quote:
> There doesn't seem to be much attention paid to SMP on Linux to date -
> MPI and other "loosely-coupled" mechanisms seem to be most popular.
> Perhaps this will change in the future.

For me, the most portable and "best" understandable mechanism for
parallel programs _is_ MPI or the like. Depending on the problem you
choose the hardware (Cray for a low communication code, SMP machine for
a high communication problem) and run the "same" code on both systems.

The only problem I see, is that the MPI implementations need a good
message passing mechanism relying on SMP possibilites, say shared memory
communication.

I just have some problems on Intel SMP hardware (Dual-PII) with Linux,
it seems, that MPICH as well as LAM/MPI are not able to use a
shmem-device as good as I expected. (Perhaps my expectations are too
high, but a (shmem-) performance improvement of less than 5% over using
a TCP/IP communication layer does not persuade me!)

So this is a Linux problem on SMP systems...!?

Perhaps someone of you made other experiences? You're welcome to prove
me wrong!

cul8r,
           Frank
--
  while (!asleep()) sheep++;



Wed, 18 Jun 1902 08:00:00 GMT  
 Parallel computing on dual-processor pentium

Quote:

> > If all you're after is using a two way SMP, then OpenMP
> > (http://www.openmp.org) may be more appropriate than
> > doing a full distributed memory message pasing re-implementation of
> > your code.

> > I believe that the PGI compiler (http://www.pgroup.com) supports OpenMP
> > code on Linux (though I haven't tried it, and there may be other compilers
> > which also do).

> I'll confirm that.

> Chose the pgroup PGI compiler for just that reason.  They've got a fairly
> decent auto-parallel code generation that can beat my pitiful attempts at
> hand coding the parallel instructions.  And if you don't like OpenMP, you
> can use their HPF compiler, too.  For a two-processor machine, I found that
> HPF tended to slow things down quite a bit compared to OpenMP.

> FWIW, for the code that I use (large matrix-vector multiplies), I get better
> results with the PGI compiler than I do with the Lahey lf95.

Thanks for the advice. This is exactly the type of code I will be using
(principally) and has been run on a Cray T3E compiled using PGHPF.  However,
comparing the pghpf compiler to the standard f90 compiler, I found that
running on (and compiling for) a single processor ,the pghpf compiler produces
code that is about 50% slower (this result still holds true for SGI Origin 2000
by the way and also compiling the code for multi-processors gives very similar
speeds when run on a single processors also) so when run on two processors I
can only get about 25% improvement on speed. I don't know if anyone else may
have similar (or opposite) experience at all.

So,  I guess my follow-up question would be, is this typical of all HPF
compilers? And, would adding MPI calls to standard F90 code be at all
deterimental to the overall performance of the code - ie can I expect better
than 25% speed increase on two processors?

Many thanks again,

Stuart.



Wed, 18 Jun 1902 08:00:00 GMT  
 Parallel computing on dual-processor pentium

Quote:
>running on (and compiling for) a single processor ,the pghpf compiler
>produces
>code that is about 50% slower (this result still holds true for SGI Origin
>2000

Yes, the hardware vendors' compilers for Mips and Alpha architectures may be
expected generally to run 50% faster with fully tuned code than alternative
compilers.  That is my experience with the R12K.  On the Intel architecture,
however, the pgi abd FSF compilers appear to be better optimized for the types
of code usually tested (including yours) than the better known commercial ones.

Tim Prince



Wed, 18 Jun 1902 08:00:00 GMT  
 Parallel computing on dual-processor pentium

Quote:

> >running on (and compiling for) a single processor ,the pghpf compiler
> >produces
> >code that is about 50% slower (this result still holds true for SGI Origin
> >2000

> Yes, the hardware vendors' compilers for Mips and Alpha architectures may be
> expected generally to run 50% faster with fully tuned code than alternative
> compilers.  That is my experience with the R12K.  On the Intel architecture,
> however, the pgi abd FSF compilers appear to be better optimized for the types
> of code usually tested (including yours) than the better known commercial ones.

However, this is misleading. The PGI HPF compiler on these platforms
is actually a pre-processor which outputs intermediate F77 (or F90)
which is then compiled with the vendor compiler.

PGI HPF therefore should benefit from all of the code-generation tweaks
in the vendor compilers.

The performance issues are more likely to be related to additional costs
introduced for handling distributed data which will lead to calls to
the run time even when only running on one processor.

-- Jim


Etnus, Inc.     +44 117 9071438
http://www.etnus.com



Wed, 18 Jun 1902 08:00:00 GMT  
 Parallel computing on dual-processor pentium

[snip]

Quote:
> There doesn't seem to be much attention paid to SMP on Linux to date -
> MPI and other "loosely-coupled" mechanisms seem to be most popular.
> Perhaps this will change in the future.

pghpf has a "-Msmp" switch that makes it produce SMP instead of MPI
code.

The current pghpf compiler does not seem to produce Fortran output any
more.  It rather looks like some intermediate format, which is then
put into a common backend.

-erik

--



Wed, 18 Jun 1902 08:00:00 GMT  
 Parallel computing on dual-processor pentium

<snip>

: comparing the pghpf compiler to the standard f90 compiler, I found that
: running on (and compiling for) a single processor ,the pghpf compiler produces
: code that is about 50% slower (this result still holds true for SGI Origin 2000

<snip>

: So,  I guess my follow-up question would be, is this typical of all HPF
: compilers? And, would adding MPI calls to standard F90 code be at all
: deterimental to the overall performance of the code - ie can I expect better
: than 25% speed increase on two processors?

I have also experienced slowdown using the pghpf compiler on SGI. This
was for a finite differencing code ported from CM Fortran, which should
map quite well into HPF. The two other HPF compilers I have used were
much better.

The free adaptor HPF compiler (converts code into PVM/MPI Fortran77/90
message passing code) seemed to give reasonable speedups on a DEC Alpha
farm (the speeds are equivelent to using PVM/MPI directly). The DEC hpf
compiler however gave the best performance, giving pretty much linear
speedup for the small number of processors available to me (1-8 cpus).

--

Mechanical Engineering,University of Sydney,NSW 2006   wk:+(61 2) 9351-2272
http://flo.mech.eng.usyd.edu.au/norris                 hm:+(61 2) 9326-5276



Wed, 18 Jun 1902 08:00:00 GMT  
 Parallel computing on dual-processor pentium
Being back at my desk after some holidays, I have resuscitated an early English draft
of my German experience report on two Fortran compilers and their parallelizers on
a dual-processor PII which appeared in the German iX journal issue 12/99 (www.ix.de/ix/).

Instead of polishing the draft literally I have inserted several screenshots which
did not appear in the final German version. Please look at
www.unics.uni-hannover.de/rrzn/gehrke/Dual-PII/fortran.html

Best wishes / WG
--
---
Wilhelm Gehrke, RRZN/UH, Schlosswenderstr. 5, D-30159 Hannover

WWW:   http://www.unics.uni-hannover.de/rrzn/gehrke/



Wed, 18 Jun 1902 08:00:00 GMT  
 Parallel computing on dual-processor pentium
On Wed, 01 Dec 1999 08:56:58 +0100, Frank Mahler

Quote:

>Well, you simply don't want to run SMP Fortran programs on NT.

You're certainly entitled to your opinion, but users of Visual KAP
have seen significant gains on a dual-processor system running NT.  I
agree that NT doesn't scale as well as good commercial UNIX systems,
but dual-processor X86 NT systems are cheap and readily available and
it seems to be a worthwhile combination for many users.

--

Fortran Engineering
Compaq Computer Corporation, Nashua NH

Compaq Fortran web site: http://www.compaq.com/fortran



Wed, 18 Jun 1902 08:00:00 GMT  
 
 [ 17 post ]  Go to page: [1] [2]

 Relevant Pages 

1. FORTRAN on dual CPU Pentium II or Pentium III

2. Parallel Computing and Distributed computing

3. Can MS SQL utilise Dual Processors?

4. Dual Processors & C5EE

5. Vendo BOARD ABIT DUAL PENTIUM

6. Clipper on NT with dual processors?

7. Clarion on Dual Processor

8. Selecting one of the dual processor on Windows

9. Oberon Support for dual Pentium II pc's

10. Dual Processors and Assembly Language

11. SIMD support on dual processor PIII?

12. How to detect Dual Processor.

 

 
Powered by phpBB® Forum Software