Subject: Re: inquiring
From: Peter Paul Smolka (smolka@uni-muenster.de)
Date: Tue Oct 20 1998 - 10:16:22 MDT
On Sun, 18 Oct 1998, Q Xu wrote:
> Date: Sun, 18 Oct 1998 20:38:31 -0600 (MDT)
> From: Q Xu <qxu@gpu.srv.ualberta.ca>
> To: ccm-users@ncar.UCAR.EDU
> Subject: inquiring
>
> Hi Friends,
>
> We are going to run ccm3. But we have not got our computation tools.
>
> Now we are in the position of either buying a Sun Ultra 10 workstation
> (512 MB memory and 300 MHZ speed) or 4 pentium PCs (128 MB memory and 300
> MHZ speed for each). If we buy PCs, we have got to install windows NT and
> some Parallel Virtual Machine (PVM) software to hook PCs to have it work as
> a single parallel computer.
>
> As we understand, 4 PCs should run 4 times faster than a single sun
> workstation because each PC processor has the same speed (300 MHZ) as the
> workstation and 4 PCs can compute at the same time. Is this concept right?
By my understanding (and experiments) you would gain under IDEAL
conditions a maximum of 2-2.5 (internal administration).
The crucial point is to "get it running at all", so check very carfully
is the compiler is able to handle the model AND is able to handle
that PVM environment.
> We do not have much idea about how to compare them. Should we buy PCs or a
> workstation?
>
Dear Dr. Xu,
my recommendation is very simple (testings made):
(1) Running ccm3.6 on NT 4.0 is possible at high speed:
1 day takes 35 minutes compared to 9 minutes mentioned for Cray in
the manual BUT:
You must have 256 MB installed (128 is NOT enough) because:
A test with a 64 MB machine took about 12 hours(!) for 1 day
(not 35 minutes). The reason was according to the taskmanager that
most of the system capacity (it was a 350 MHz PII) was used for swapping,
the CPU was overall running at about5-10 percent sometimes up to 80
(short peaks)).
All tests have been made at T42 resolution.
2) So:
If you buy ONE Pentium II/300 with 256 MB RAM and NT 4.0 you
should arrive at above 35 minues per day.
You should use DEC Visual Fortan V5 and look carefully at all
passages dealing with the file-system, especially program-generated
filenames. The other system-dependent issues (date/time etc) are
"minor technical rubbish".
If AFTER such an experiment (3000 USD) you are still unsatisfied with
35 minutes,buying expensive SUNs is possible anyway. And NOTE:
It is really the memory. The model itsself needs 167 MB and with
NT together it is 207.
> To hook PCs needs knowledge of network. Is it difficult if any problem
> happens to the network? Message transfer speed should also be a major
> factor. Do we need any special software to make a good speed of message
> transfer in the hooked PCs?
>
I mean: For above computational speed some network acrobatics is not
necessary, even a dual processor system is not necessary but just:
console-mode (no gui)and shere processor power).
Auction-Sales.com offered some months ago even an ALTO Pentium II/800
so in case of dissatisfaction with above 35 minutes some dollars more
bring you much faster to your results then complex message-passing
thoughts. (I am definitelyNOT against parallel architecture, we had
transputers here even since 1988, but despite the "technical sweetness"
the tradeoff in terms of results/(time*cost) is too low).
> How about running time of 10-year simulation period using T42 resolution on
> both sun ultra-10 workstation and hooked PCs?
10 years not yet tried, but multiplying 35 minutes for a day by 3650
should be the pessismistic result as the first 5 minutes (about) are
used for initial settings.
I mean: Keeping the PC running for some 80 days needs some patience...
BUT: Make first(!) a test n with the Sun. According to our experience
their speed in the public face is greater than in the machine (I do
believe in benchmarks, but we do not need benchmarks but
overall performance). So: Ask a person who runs ccm3.6 on a SUN for the
time it needs for one day.
>
> Could you also tell me the version of C and Fortran compiler to run ccm3?
>
As said above: Dec Fortran V5. In case of compiler options we can give
you hints.
All works testings etc done as part of UNESCO IGCP-341, Southern Hemisphere
Paleo and Neoclimates.
> Any suggestions and advices are welcome!
>
So overall my stragey would be (if you have no deadlines of the financial
year):
Buy 1 fast PC with 100 MHz board, 256 MB RAM, single processor, 6 GB
IDE disk (you need 50-200 MB). Processor: 300 or 400 MHz.
Get it running and make experiments over weeks, month and a year.
(which means: USD 3000)
THEN buy after 6-9 months the fastes PC on the market (some Alto Systems
800 MHz Pentium II with fast board (we have the "old" 67 Mhz board))
and cut the computational speed by 2-2.5 (or more if the Gigahertz
iron is available). This strategy leaves you sufficient money for
experiments in any directions and before buying a SUN ask some persons
for the minutes/day. Just "let the numbers speak".
And: Do NOT economize on memory.
> Yours sincerely,
>
Best regards
Peter Smolka
Dr. Peter P. Smolka
University Muenster
Geological Institute
Corrensstr. 24
D-48149 Muenster
Tel.: +49/251/833-3989 +49/2533/4401
Fax: +49/251/833-3968 +49/2433/4401
Try to leave the earth as you would like to meet her when you come.
This archive was generated by hypermail 2b27 : Thu Jun 01 2000 - 09:07:08 MDT