Re: CCL:Summary: AMD-K6/linux for G98



 On Wed, 24 Feb 1999, Jerry C.C. Chan wrote:
 >
 > Sorry for any duplication, it seems the following summary didn't get
 > posted at the first time.
 > -------------------------
 >
 > Many thanks for those who responsed to my post.  In short, P-II is highly
 > recommended, DEC Alpha (21164) is a good alternative, Celeron-based
 > machine MAY be a good choice (see both positive and negative comments
 > below), AMD-K6 is not at all useful for numerical calculation.  The 11th
 > response from Prof Teppen is very interesting that the performance of his
 > 4-processor-P-II linux box from Peter Pulay's company (cost $11,000) is
 > only 15% slower than a Cray T3E-1200 with 4 processors.  Please excuse me
 > for removing some of the signatures in order to shorten this summary.
 >
 > Original post:
 > > We have a plan to get a linux box equipped with AMD-K6 for Gaussian 98
 > > calculations.  However, it seems that the recommended blas library and
 > > Portland Fortran compiler are optimized for Intel CPU only (I may be
 > > mistaken saying so).  I would be grateful if anyone could share
 his/her
 > > experience concerning the following thoughts
 > > i. pay more money to buy Intel processor
 > > ii. go on with our plan (assuming that the performance of
 Linux/AMDK6-G98
 > > won't be very different from Linux/IntelPII-G98)
 > > iii. keep the AMD but switch to NT/G98W instead of linux/G98
 >
 > ++++++Responses++++++
 > 1. From: Drake Diedrich <Drake.Diedrich (- at -) anu.edu.au>
 >
 >    BLAS isn't apparently heavily used in Gaussian (only in CI methods),
 > but...  you may still be better off getting an Intel CPU.  On heavy
 floating
 > point and memory bandwidth I have benchmarked a pentium-166 outperforming
 an
 > AMD K6/2-300. Pentiums (and later Intel processors) have pipelined floating
 > point units while AMD and Cyrix have only a single floating point execution
 > unit.  AMD processors are not competitive with the pentium-II.
 >    You should add the DEC Alpha (21164) to your shortlist of alternatives.
 > I haven't benchmarked Gaussian on it, but if you search the web you'll find
 > a hand coded BLAS library for the Alpha that achives 800 MFlops on a 600MHz
 > chip.  The ASCI-Red BLAS library gets only 180 MFlops on our
 Pentium-II/266.
 >
 > -Drake
 >
 > 2. From fparnold (- at -) balihai.uchicago.edu Tue Feb 23 10:46:42 1999
 >
 > If you follow the links through the Beowulf page (www.beowulf.org) to
 > their archive for February of 1999 (or possible Jan) there is a comparison
 > of CHARMm run times on the K6 vs. Celeron.  Basically, CHARMm is slow on
 > the K6, but runs on teh Celeron at the same speed of the PII.  So, if
 > you're not cache limited, this is a cheap option.  IF you are, then the K6
 > if probably ok.  My personal results with GAMESS-US indicates that the
 > K6-233 is slightly faster than a P200 classic, and roughly 5% faster than
 > an SGI O2 with an R5000 processor and secondary cache.
 >
 > Unfortunately, the real answer therefore is, "buy one, and benchmark
 it
 > yourself".  If you get good numbers, buy more and make a cluster out
 of
 > them.  If you don't, you can probably return it and just be out the
 > shipping fee.
 > 							-fred
 >
 > 3. From pczyryca (- at -) smurf.chem.usu.edu Tue Feb 23 10:46:45 1999
 >
 > K6 is slower than PentiumII (with the same clock). Especially on floating
 > point operations. This is not OS dependent, so I wouldn't expect better
 > performance with NT/G98W combination.
 > Probably it's the best option to spend some more money and buy PentiumII.
 > Regards,
 >
 >  * Przemek Czyryca, pczyryca (- at -) smurf.chem.usu.edu *
 >
 > 4. From gammadas (- at -) telis.org Tue Feb 23 10:46:49 1999
 >
 > I had the same questions and similar experiences  for g94w/g98w as
 > mentioned in the  message below---I would pay more and get the Intel
 > PII(PIII will be available in 2 wks!)  Also check 2nd message (Jerry: the
 > next message).  Hope it helps!
 >
 > 5. From: "Dr.Christian M¸ck-Lichtenfeld" <cml (- at -)
 lipari.uni-muenster.de>
 >
 > I have a small number of PCs at hand, which are run under
 > Windows NT and Linux. Although the hardware is not exactly
 > the same, one can come to the overall conclusion that the
 > AMD processors are about 1.5 - 2 times slower than the
 > Intel machines. Even a PentiumII with 266 MHz was found to
 > run a moderate G98W-job in about 60% of the time that an
 > AMD K6/2 with 300 MHz needed (both with at least 128MB RAM).
 > A comparison of several Mopac jobs (Mopac93/Linux) on the
 > AMD K6/2 (300MHz) and on a PentiumII (333MHz) gave the same
 > result (t(AMD)/t(PII) = 1.5 ... 2).
 >
 > I am not a hardware expert but I would suppose that the different
 > architecture of the CPUs clearly favours Intel over AMD, due to the better
 > performance in floating point operations.  For the aquisition of new
 > machines I therefore would strongly suggest Intel over AMD.
 >
 > I would be very interested to hear from the experiences of other
 > users, so please forward their answers or summarize if possible.
 >
 > Best wishes
 >
 > Christian Mueck-Lichtenfeld
 >
 > 7. From Steven.Creve (- at -) dsm-group.com Tue Feb 23 10:46:52 1999
 >
 > do anything but the last suggestion. WindowsNT is really a *bad* choice for
 > doing (any kind of) calculations.
 > I would suggest you contact the people at Gaussian with this problem.
 > Steven
 >
 > 8. From: val <val (- at -) cacr.ioc.ac.ru>
 >
 > > i. pay more money to buy Intel processor
 >
 >   It seems yes, see comments below.
 >
 > > ii. go on with our plan (assuming that the performance of
 Linux/AMDK6-G98
 > > won't be very different from Linux/IntelPII-G98)
 >
 >   Using GAMESS ab-initio program under Linux we found that Intel PII-350
 >   is  significantly faster then AMDK6-350 (in order of 50%) even without
 >   any specific blas libraries. If your budget allows, go to the PII CPU.
 >
 >   On the other hand, Celeron shows very good performance: Celeron-300A is
 >   close to AMDK6-350, Celeron 333 expected to be a little faster. Both are
 >   really cheap now.
 >
 >   I have a small collection of GAMESS/ab-initio benchmarks under linux on
 >   my homepage:    http://nmr.ioc.ac.ru/Staff/AnanikovVP/
 >
 > > iii. keep the AMD but switch to NT/G98W instead of linux/G98
 >
 >   This is very old (and dangerous ;) question, you may look to the CCL
 >   archives, the last discussion finished not long ago.
 >
 >   For example, a part of the email from Mike Frisch <frisch (- at -)
 lorentzian.com>;
 >   CCL Tue, 5 Jan 1999:
 > ".. We use the same compiler (Portand) and blas libraries on both
 Windows/NT
 >     and Linux, and performance is similar.  The differences in reported
 times
 >     are a few percent and probably reflect Linux reporting system time
 >     overhead (as other unix versions do) while this isn't included on a
 >     per-process basis under NT.  Elapsed times are sometimes a bit better
 >     under Linux because of better I/O, but the best summary of the
 situation
 >     is that there isn't a signficant performance diffence for running a
 >     single, stand-alone job under one OS or the other."
 >
 > best regards,
 > Valentin.
 >
 > 9. From ivan (- at -) lipid.biocomp.unibo.it Tue Feb 23 10:46:57 1999
 >
 > > i. pay more money to buy Intel processor
 >
 > That is the way to go. AMD K6 is a great cpu, but NOT for scientific and
 > technical applications. Its Floating-point unit is VERY slow when compared
 > to intel one. Get a real PII, the fastest you can afford, don't buy a
 > Celeron-based machine.
 >
 > BTW i tested both CPU with G94.
 >
 > > ii. go on with our plan (assuming that the performance of
 Linux/AMDK6-G98
 > > won't be very different from Linux/IntelPII-G98)
 >
 > Once more, just to stress it....
 >
 > I don't think the small saving is absolutely worth the the performance
 > penalty that you have with K6 with scientific and technical programs.
 >
 >
 > > iii. keep the AMD but switch to NT/G98W instead of linux/G98
 >
 > linux is more stable.  go for it.
 >
 > --
 > Dr. Ivan Rossi - CIRB Biocomputing Unit
 > e-mail: ivan (- at -) lipid.biocomp.unibo.it  WWW: http://www.biocomp.unibo.it/ivan
 >
 > 10. From daniels (- at -) kemi.uu.se Tue Feb 23 10:47:00 1999
 >
 > I haven't tried gaussian on the AMD-k6, but from my experiences I would
 *not*
 > recommend using the K6 for floating point stuff. I really like the
 performance
 > of the K6 for regular stuff like editing, compiling, but the fp perfomance
 is
 > simply not good enough. my experience is that the pii is about 50 % faster
 than
 > the k6 with the same frequency. the processor is only part of the computer
 > anyway, you still have to pay the same money for case, monitor, disks etc,
 so
 > if the final computer is perhaps 20 % more expensive with the PII, it is
 still
 > worth it for fp stuff. What may break down my argument is that in my
 > calculations, the fp is the bottleneck. In your case it may be the disk,
 so...
 >
 > best regards
 > Daniel Spångberg
 > Inorganic Chemistry
 > Uppsala University
 > Sweden.
 >
 > 11. From teppen (- at -) pilot.msu.edu Tue Feb 23 10:47:05 1999
 >
 > If I may propose a fourth option in addition to the three you posted:
 >
 > Peter Pulay at the University of Arkansas is a true pioneer of quantum
 > chemistry - for example, his contributions were cited in Pople's official
 > Nobel Prize announcement this past fall. He is a former professor of mine
 > and I like him very much, but I have no other interest in promoting his new
 > company.
 >
 > Pulay has been working on optimized quantum codes since the 1960s and he
 > just started a company to sell them this past year. He has 5-6 coworkers
 > including Jon Baker, another master of optimization methods. Several years
 > ago, Pulay began to parallelize his program for cheap PCs running Linux,
 > and they now have a system they can sell. Their company is "Parallel
 > Quantum Solutions" and pulay's e-mail is "pulay (- at -)
 pqs-chem.com". They build
 > their own PCs and sell a software-hardware bundle.
 >
 > I paid $11,000 for a 4-processor (400 MHz Pentium IIs) machine with Linux
 > and the quantum software installed. I have just got the machine, but
 > benchmarks indicate that this machine has exceedingly good
 > performance/price. For pinene using DFT/B3LYP with large basis sets, for
 > example, the machine i bought for $11,000 was faster than a Cray T3E-600
 > with 4 processors and only 15% slower than a Cray T3E-1200 with 4
 > processors, both running parallel Gaussian 94. (The G94 calculations were
 > done by Sosa and Frisch, so were probably pretty well optimized for the
 > machine, see last July 15 number of J. Comput. Chem. Sosa et al. Vol 19, p.
 > 1053). I haven't priced those machines but they must be >$100k. If you
 want
 > to benchmark against one of your own machines, Pulay would be happy to do
 > some calculations for you.
 >
 > Again, I have no financial interest in telling you this, but it seems like
 > a great product for working quantum chemists. The graphical interface is
 > not great, but it is adequate. They put their time into optimizing code.
 >
 > Best wishes,
 >
 > Brian
 >
 > **********************
 > Dr. Brian J. Teppen
 > Assistant Professor
 > Department of Crop and Soil Sciences
 > Plant and Soil Science Building, Room 538
 > Michigan State University
 > East Lansing, MI 48824-1325
 >
 > phone: 517-353-5215
 > fax: 517-355-0270
 > e-mail: teppen (- at -) pilot.msu.edu
 >
 > 12. From haasj (- at -) ubaclu.unibas.ch Tue Feb 23 10:47:08 1999
 >
 > i am a chemistry student in the very last three weeks of studying. I am
 > using g98 /molden /k6-2-300 and RH 5.2 Linux and 96MB of RAM. I have
 > encountered that the speed is ok, but not as fast compared to a PII. If i
 > might I would recon you to buy the K7 (will probably come out this year)
 > with the EV6 bus protocol of a DEC ALPHA processor 21264. In later times
 > you would have the opportunity to exchange the proc against an 21264 -
 > what ever clockspeed... nice option i think. With a PII you are at the
 > limit of only increasing clockspeed if even possible, although PII are
 > better in floating point operations. DEC Alpha is for sure a good choice
 > for later. K6-3 (sharptooth) is also quite nice proc to think of cheaper
 > than PII but should be equally fast.. (I can only assume that NT system is
 > equally fast)...
 >
 > greetings juergen haas
 > uni basel
 > org. chem dep.
 >
 >
 > feel free to contact me again
 >
 > 13. From eugene.leitl (- at -) lrz.uni-muenchen.de Tue Feb 23 10:47:10 1999
 >
 > AMD K6 ist voll kompatibel zu PPro/P II CPU. Allerdings koennte
 > Celeron das bessere Preis/Leistungsverhaeltnis bieten. Vielleicht auf
 > K6-3 oder sogar K7 warten?
 >
 > MfG,
 > Eugen
 >
 > 14. From RedAndr (- at -) anrb.ru Tue Feb 23 10:47:12 1999
 >
 > It is not good idea buying AMD/K6 CPU for floating point calculations.
 > Intel processors with same price as AMD have more powerful FPU
 > performance. In addition all compilers can do optimization only for
 > Pentium/Pentium Pro unlike AMD or Cyrix. Therefore you have to forget
 > about using AMD in numerical purposes. You can also buy multiprocessor
 > station with Intel processors unlike AMD. About Linux and NT. Our
 > experiments shows that Linux haven't advantage over Windows 95/NT in
 > numerical calculations.
 >
 >
 > Best regards,
 >       Andrew                            mailto:RedAndr (- at -)
 anrb.ru
 >
 >
 > 15. From gaussian.com!fox (- at -) lorentzian.com Tue Feb 23 10:47:15 1999
 >
 >     At this point I cannot vouch for proper operation of G98 on an
 > AMD processor.  We do not have plans for supporting this as a platform.
 > While most things work on AMD I have at least one site with a problem
 > trying to run G98W and have several sites that have demonstrated problems
 > with Cyrix processors which make it impossible to run G98.
 >
 >    In short I recommend you get Intel.
 >
 >   Douglas J. Fox
 >   Director of Technical Support
 >   help (- at -) gaussian.com
 >
 > -----end of summary------
 >
 >
 >
 >
 >
 >
 > ---
 > Administrivia: This message is automatically appended by the mail exploder:
 > CHEMISTRY (- at -) www.ccl.net: Everybody | CHEMISTRY-REQUEST (- at -)
 www.ccl.net: Coordinator
 > MAILSERV (- at -) www.ccl.net: HELP CHEMISTRY or HELP SEARCH | Gopher:
 www.ccl.net 73
 > Anon. ftp: www.ccl.net   | CHEMISTRY-SEARCH (- at -) www.ccl.net -- archive
 search
 >              Web: http://www.ccl.net/chemistry.html
 > ---
 >
 >