Need for speed.

  • Thread starter Thread starter Russel Sprout
  • Start date Start date
R

Russel Sprout

Well not extreme speed, but a little better than I have now - AMD Athlon XP
2200+.

I am not interested in gaming or video/image manipulation, just pure umber
crunching. I'm working on a project where a program is run several times,
currently each run takes around 10 mins, CPU is running at 100%, but its not
using much RAM (PC2100) and disk activity seems to be low. I'd like to cut
this to 2 mins or below. Specifically I'm running degrib with the -prb
swich, if that means anything to anyone. It decodes compressed weather model
data in GRIB2 format, files are typically big, 32Mb.

Any sugestions, price is a consideration, but I'd like the best bang for
bucks. The program is 32 bit, but I think there may be a 64 bit version and
it may also be optimised for multiple proccesers.

TIA
Steve
 
Russel said:
I am not interested in gaming or video/image manipulation, just pure umber
crunching. I'm working on a project where a program is run several times,
currently each run takes around 10 mins, CPU is running at 100%, but its not
using much RAM (PC2100) and disk activity seems to be low. I'd like to cut
this to 2 mins or below. Specifically I'm running degrib with the -prb
swich, if that means anything to anyone. It decodes compressed weather model
data in GRIB2 format, files are typically big, 32Mb.

What type of numbers are being calculated? Int32, Int64, 64-bit Float,
80-bit Float or any of the Streaming (SIMD/SSE) calculations?
 
Russel said:
Well not extreme speed, but a little better than I have now - AMD Athlon XP
2200+.

I am not interested in gaming or video/image manipulation, just pure umber
crunching. I'm working on a project where a program is run several times,
currently each run takes around 10 mins, CPU is running at 100%, but its not
using much RAM (PC2100) and disk activity seems to be low. I'd like to cut
this to 2 mins or below. Specifically I'm running degrib with the -prb
swich, if that means anything to anyone. It decodes compressed weather model
data in GRIB2 format, files are typically big, 32Mb.

Any sugestions, price is a consideration, but I'd like the best bang for
bucks. The program is 32 bit, but I think there may be a 64 bit version and
it may also be optimised for multiple processors.

It seems to me degrib is reference software.
http://www.nws.noaa.gov/mdl/NDFD_GRIB2Decoder/

This could mean is has *NOT* been optimized.

Did you try to make your own build?
If so, did you pass any optimization flags to the compiler?
 
gaf1234567890 said:
What type of numbers are being calculated? Int32, Int64, 64-bit Float,
80-bit Float or any of the Streaming (SIMD/SSE) calculations?

I doubt there is any need for 80 bit float. I'm going to try compiling
myself as Grumble suggests, so I'll take a look at the code. I'll also spend
a few hundred on a faster processor and better RAM and HDD. I'll probably
build the machine myself as most of the higher end PCs seem to have flashy
video/sound/TV cards etc which ! will not needed and then they seem to skimp
a little on HDDs.

Any further suggestions on a good MOBO, CPU and RAM conbination will be
appreciated.

Thanks all.

Steve.
 
I doubt there is any need for 80 bit float. I'm going to try compiling
myself as Grumble suggests, so I'll take a look at the code. I'll also spend
a few hundred on a faster processor and better RAM and HDD. I'll probably
build the machine myself as most of the higher end PCs seem to have flashy
video/sound/TV cards etc which ! will not needed and then they seem to skimp
a little on HDDs.

Any further suggestions on a good MOBO, CPU and RAM conbination will be
appreciated.

Since you don't care about gaming, try a mobo with integrated video.
Recently I had a good experience with MSI RS482M4 - nice little
micro-ATX with surprizingly good video (for integrated one, that is)
and even dual video output - one DVI, one VGA, $74 @ Newegg.com. NIC
and sound (hi-def) included. RAM - up to 4 GB DDR 3200 (aka pc400),
but your old 2100 will work. CPU - any A64 939 that fits your budget,
from 3000 all the way up to FX60.

Good luck.

NNN
 
I doubt there is any need for 80 bit float. I'm going to try compiling
myself as Grumble suggests, so I'll take a look at the code. I'll also spend
a few hundred on a faster processor and better RAM and HDD. I'll probably
build the machine myself as most of the higher end PCs seem to have flashy
video/sound/TV cards etc which ! will not needed and then they seem to skimp
a little on HDDs.

Any further suggestions on a good MOBO, CPU and RAM conbination will be
appreciated.

First, you need to figure if a dual core processor is worth anything to
you, either because the program is efficiently multi-threaded or possibly
because you want to be able to do other useful things while it's running.

I don't know your application but for most general purpose work AMD64 CPUs
are beating Intel P4s quite handily and running cooler into the bargain.
Next you need to figure if ECC memory is desirable. The AMD64 field is a
bit confused right now: there are Athlon64s & Opterons which fit in a
socket 939 and Opterons which fit in a socket 940; both Athlon64s &
Opterons can do ECC but AFAIK there is only one mbrd mfr, Tyan, who makes a
s939 version. All s940 mbrds should handle ECC.

If you go s939, it doesn't matter if you get an Opteron or Athlon64 - the
former is claimed by the overclocking/gamer crowd to have a better memory
controller but I'm not sure why. In fact I'm baffled as to why AMD has
decided to sell s939 Opterons - it only makes choosing a CPU confusing.

For a mbrd, I've been using MSI recently and I haven't had any problems
with them apart from some which came with a bad batch of chipset fans...
which they replaced promptly. MSI's BIOS updates, for new CPU
compatibility and chipset feature improvements have been very good and
timely. Now you need to figure which chipset feature set you want: e.g. is
Firewire and dual network interfaces something you'd want?... things which
are generally only on the higher priced mbrds.

For the mbrd chipset, I've used both VIA and nVidia and for performance
it's a toss-up - nVidia has been a bit in front with latest features like
SATA-2 but the advantage is debatable. nVidia has some fancy hardware
assist in their NIs but it just doesn't work right - in fact I had to turn
off Checksum Offload to get the damned thing to work right.

For memory, I wouldn't even bother considering anything but Crucial but for
best performance, you want to keep the rank count down: with AMD64 CPUs,
above 3 ranks of memory, you might have to back off on memory timings or
clock speed. This means you should try to get DIMMs with 512M-bit chips...
preferably with the Micron memory chips; Crucial *is* selling some DIMMs
just now with Samsung chips and they are fussier about timing.

For a HDD, especially SATA-2, Seagate is working well (best ?) with nVidia
nForce4 chipsets. Maxtor has, of course been taken over by Seagate
recently but there have been problems with their HDDs at SATA-2 speeds.

If you tell a bit more about features important to you you'll get pointers
to specific mbrds.

I'm not sure you're going to get down to 2mins for your jobs but I'd think
5mins is certainly feasible. Good luck with it.
 
Well not extreme speed, but a little better than I have now - AMD Athlon XP
2200+.

I am not interested in gaming or video/image manipulation, just pure umber
crunching. I'm working on a project where a program is run several times,
currently each run takes around 10 mins, CPU is running at 100%, but its not
using much RAM (PC2100) and disk activity seems to be low. I'd like to cut
this to 2 mins or below. Specifically I'm running degrib with the -prb
swich, if that means anything to anyone. It decodes compressed weather model
data in GRIB2 format, files are typically big, 32Mb.

Any sugestions, price is a consideration, but I'd like the best bang for
bucks. The program is 32 bit, but I think there may be a 64 bit version and
it may also be optimised for multiple proccesers.

Well, I don't know much about the GRIB2 or the 'degrib' program
myself, but from your description and the link that Grumble provided,
my guess is that it's going to have somewhat similar performance
characteristics to something like WinRAR or gzip, though perhaps with
less disk activity (or maybe not.. tough to say). The processing
requirements seem to revolve around unpacking and sorting data. Given
that I don't see any actual 'degrib' benchmarks, perhaps some WinRAR,
WinZIP or gzip tests might give a bit of insight. Try the following
links:

http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2484&p=6

http://www.neoseeker.com/Articles/Hardware/Reviews/a64-x2-4200/6.html

http://www.pcstats.com/articleview.cfm?articleid=1918&page=15

And if you're really interesting:

http://www.spec.org/cpu2000/results/cint2000.html

164.gzip is one of the subtests of SPEC CINT2000 and has been tested
on nearly every CPU out there.


Note that almost none of the above are multithreaded. If you're
'degrib' program is indeed multithreaded (ie able to work with
multiple processors) then it will likely see a SIGNIFICANTLY
improvement with a newer dual-core chip. I would guess that an
application like this could probably see at least a 60% boost in
performance when going for a single core processor to an equivalent
dual-core processor, and possibly over a 90% improvement. Even if the
program is not multithreaded you may want to consider a dual-core
processor since it will give you a free CPU core to do other work
(even if it's just checking e-mail and such) while 'degrib' is running
on the first core.
 
Russel Sprout said:
Well not extreme speed, but a little better than I have now - AMD Athlon XP
2200+.

I am not interested in gaming or video/image manipulation, just pure umber
crunching. I'm working on a project where a program is run several times,
currently each run takes around 10 mins, CPU is running at 100%, but its not
using much RAM (PC2100) and disk activity seems to be low. I'd like to cut
this to 2 mins or below. Specifically I'm running degrib with the -prb
swich, if that means anything to anyone. It decodes compressed weather model
data in GRIB2 format, files are typically big, 32Mb.

Any sugestions, price is a consideration, but I'd like the best bang for
bucks. The program is 32 bit, but I think there may be a 64 bit version and
it may also be optimised for multiple proccesers.

TIA
Steve
Thanks for all the replies, lots to think about!

Steve
 
George said:
I don't know your application but for most general purpose work AMD64
CPUs are beating Intel P4s quite handily and running cooler into the
bargain. Next you need to figure if ECC memory is desirable. The
AMD64 field is a bit confused right now: there are Athlon64s &
Opterons which fit in a socket 939 and Opterons which fit in a socket
940; both Athlon64s & Opterons can do ECC but AFAIK there is only one
mbrd mfr, Tyan, who makes a s939 version. All s940 mbrds should
handle ECC.

Unless I am mistaken, most socket 939 motherboards do accept
unbuffered ECC DIMMs.

e.g. my ASUS A8N-E's spec says:

4 x 184-pin DIMM Sockets support max. 4GB DDR400/DDR333/DDR266
ECC/non-ECC un-buffered DDR SDRAM memory

and Crucial's Memory Advisor suggests several ECC modules.

http://www.crucial.com/store/listparts.asp?model=A8N-E&tabid=AM
 
Unless I am mistaken, most socket 939 motherboards do accept
unbuffered ECC DIMMs.

Thanks for that correction - MSI's mbrds don't support ECC but apparently
Asus and some others may.
e.g. my ASUS A8N-E's spec says:

4 x 184-pin DIMM Sockets support max. 4GB DDR400/DDR333/DDR266
ECC/non-ECC un-buffered DDR SDRAM memory

Yes OK *but* according to Asus' qualified vendor list there are very few
ECC DIMMs which work in their s939s... possibly because it's a market --
i.e. unbuffered ECC -- the enthusiast memory providers don't address.
 
..
Thanks for all the replies, lots to think about!

Steve


Does anyone know if AMD markets their own compiler? I was working with some
guys that were doing some intensive math work back a year or so ago on
RedHat ES, They had found that Intel markets a compiler for Linux and it
really helped out for some of the calculations they needed to do.
 
.


Does anyone know if AMD markets their own compiler? I was working with some
guys that were doing some intensive math work back a year or so ago on
RedHat ES, They had found that Intel markets a compiler for Linux and it
really helped out for some of the calculations they needed to do.

Which language?;-) Intel has marketed various compiler
enhancements/plugins over the years, some of which had a dreadful
reputation for stability/usability for making "production" code. They were
designed for primarily for err, benchmarking. I believe they inherited the
Digital Visual Fortran Compiler and it had a good reputation for code
quality and speed... no idea what it's current status is - no idea what
they have in C/C++ currently either. Serious developers got very wary of
Intel's compilers a while back... but that may have changed more recently.
Then again you *know* that Intel's compiler is going to be umm, trained for
Intel CPUs and their quirks... why bother?

AMD does not market a compiler but offers advice on switches/usage of
others' compilers etc. here http://developer.amd.com/Default.aspx.
 
Which language?;-) Intel has marketed various compiler
enhancements/plugins over the years, some of which had a dreadful
reputation for stability/usability for making "production" code. They
were designed for primarily for err, benchmarking. I believe they
inherited the Digital Visual Fortran Compiler and it had a good
reputation for code quality and speed... no idea what it's current
status is - no idea what they have in C/C++ currently either. Serious
developers got very wary of Intel's compilers a while back... but that
may have changed more recently. Then again you *know* that Intel's
compiler is going to be umm, trained for Intel CPUs and their
quirks... why bother?

AMD does not market a compiler but offers advice on switches/usage of
others' compilers etc. here http://developer.amd.com/Default.aspx.


In this case it was a mix of C and C++, or should I say, C written in kinda
C++. I never gave Intel compilers much of a chance but I was impressed
with the speed increase it gave for this case. For all the code of I been
writing recently, I don't even care much most of the stuff my group works
on is non processor intensive.

I think this was what they were using.
<http://www.intel.com/cd/software/products/asmo-
na/eng/compilers/clin/277618.htm>
 
Back
Top