Summary:Benchmarks
Dear Netters,
Some days ago I posted a message about new HP and SGI workstations perfomance.
My original question and answers are shown below.
The last part of this message is a review by Martyn F. Guest at CCLRC Daresbury
Laboratory about 'Performance of Various Computers in Computational Chemistry'
that includes lots of useful data.
Thanks to all who answered:
Steen Hammerum
Bill DeSimone
Jose Luis Garcia de Paz
Oakley H. Crawford
Marc C. Nicklaus
Vikram Varma
Alexander Hofmann
Martyn F. Guest
========= ========= ========= ======== ****** ******
From: Steen Hammerum <steen -8 at 8- kiku.dk>
I recently asked the very same question, and here is the summary I posted:
------------------------------------------------------------------------
Summary: Speed of new HP and SGI workstations
I recently asked this group for advice with regard to the performance
of new HP and SGI workstations running the Gaussian programs
------------------------------------------------------------------------
Does anyone have or know of benchmarks for Gaussian 94 or related programs
running on the new SGI and HP chips (R10000 and PA8000), or other
information to assist us before we decide on new machinery?
We are currently considering SGI Power Indigo 2 and HP C160 workstations
that will be used predominantly for Gaussian calculations, but we are not
sure how well the SPECint95 and SPECfp95 numbers allow us to assess the
relative performance.
-------------------------------------------------------------------------
---
Roberto Gomperts (roberto -8 at 8- boston.sgi.com) provided the following
comparison of the SGI R8000 and R10000 chips when running test178
from the Gaussian 94 test suite (a single point, direct scf calculation
with 300 basis functions), using Gaussian 94, rev. D3.
Machine Chip/Frequency Sec. Cache Time(min.)
Power Indigo2 r8k/75 MHz 2 MB 8.04
Power Challenge r8k/90 MHz 4 MB 6.50
Power Indigo2 r10k/195 MHz 1 MB 7.30
Power Challenge r10k/195 MHz 1 MB 7.08
Power Challenge r10k/195 MHz 2 MB 5.95
The Power Challenge runs are done on 1 processor. These are all CPU times.
The Wall clock times are very similar to these (less than 20 sec. difference)
---
John Brodholt (j.brodholt -8 at 8- ucl.ac.uk) pointed me to
http://gserv1.dl.ac.uk/TCSC/disco/TechPapers/bench/bench.html
This site provides a detailed, critical and very useful
comparison of the performance of a wide variety
of newer workstations; unfortunately, no results obtained with HP PA8000
machines are included (html version not yet available, but a postscript
version can be downloaded).
The comparison is based on timing data obtained with GAMESS-UK
(rather than Gaussian). The results illustrate that the relative
performance varies quite a bit with the type of calculation undertaken.
---
Eric Billings (billings -8 at 8- helix.nih.gov) pointed me to
http://www.ki.si/parallel/summary.html
The information was designed to compare parallel architectures, but the
single CPU column provides useful information.
---
Finally, Glenn McEnroe (gmcenroe -8 at 8- crl.com) suggested that I looked
elsewhere:
"Regarding your question about SGI vs HP you should check out the latest
issue of Journal of Computational Chem V17 No. 11 1385-86 entitled
Viability of Molecular Modeling with Pentium based PCs. This article does
not compare these new chips for SGI and HP but it appears that you may be
better off running your application on a pentium based machine if Gaussian
94 is available for this platform."
---
Many thanks to everyone who answered.
----------------------------------------------------------------
Since writing the summary, I have had the opportunity to perform trisl
calculations on both new SGI and new HP machines. The results with R10k
SGI machines confirm Roberto Gomperts' results (see summary), that is, the
new chip is very fast but not much faster than the old chip, and cache size
matters a lot. My HP results have been somewhat disappointing, insofar as
the new chip in real life situations (if G94 calculations can be called
"real life") does not seem to be quite as fast as the benchmarks would
lead
you to expect.
Hope this helps,
Steen
--
Steen Hammerum steen -8 at 8- kiku.dk
Department of Chemistry (+45) 35 32 02 08
University of Copenhagen, Denmark fax: (+45) 35 32 02 12
========= ========= ========= ======== ****** ******
From: desimone -8 at 8- mroa.ENET.dec.com
Carles,
The AlphaStation 500 with the 500 MHz chip should perform at least
as well if not better than HP and SGI on Gaussian and AMBER. A
customer in Isreal just bought an AlphaStation 500 to run Gaussian.
I can see if this customer would talk to you if you are interested.
Unfortunately all my Gaussian benchmarks are on AlphaServers ,
which BTW, beat the competition on Gaussian benchmarks.
Also, the UCSF data on AMBER is on very old AlphaStations.
Bill DeSimone
Science & Research Applications
High Performance Computing
DEC
========= ========= ========= ======== ****** ******
From: DEPAZ -8 at 8- ccuam3.sdi.uam.es
Carles,
En la Univ Autonoma de Madrid se acaba de comprar un DEC de 8 cpu, en
competencia con un SGI de ocho r10000. Hicimos pruebas con el
gaussian (en eso estuve yo) y EN TODAS nos salio un 20% mas
repido el chip dec que el chip silicon.
Hubo test del gaussian usando una sola cpu, usando varias (paralelo), etc.
No hubo color. La version gaussian para dec iba mejor que la gaussian
que habian modificado para silicon.
Yo estuve en la comision tecnica como quimico cuantico y no tuve dudas.
Un saludo
Jose Luis Garcia de Paz
quimica fisica aplicada
3974263-4957
========= ========= ========= ======== ****** ******
From:crawfordoh -8 at 8- ornl.gov
If you have access to the www, look at
http://www.netlib.org/performance/html/PDStop.html
for performance data on a vast array of machines.
Otherwise, send the following message to netlib -8 at 8- ornl.gov, to receive
performance data in postscript files:
send performance.ps from benchmark
send mp-computers.ps from benchmark
Finally, if you want more info about netlib's collection, send the following
message:
send index
Good luck,
Oakley Crawford
----------------------------------------------------------------------------
Oakley H. Crawford Phone: +1-423-574-5048
Oak Ridge National Laboratory Fax: +1-423-574-6210
P. O. Box 2008, MS 6123 E-mail: crawfordoh -8 at 8- ornl.gov
Oak Ridge, Tennessee 37831-6123 Express delivery, add: Bethel Valley Road
USA
----------------------------------------------------------------------------
========= ========= ========= ======== ****** ******
From: "M. Nicklaus" <mn1 -8 at 8- helix.nih.gov>
Dear Dr. Colominas,
We have a four-processor Digital AlphaServer 4/275 which we use mostly
to run Gaussian 94 and the Molecular Mechanics program CHARMM. We are
very happy with it. We have not done benchmark comparisons with an SG
R10000 system ourselves (since we don't have one), but I remember having
seen quite a few postings of, or at least pointers to, benchmarks with
ab initio and MM programs that included SGI, HP, and/or DEC systems, and
which should be retrievable from the CCL archives. Hope this helps.
Regards,
Marc C. Nicklaus
------------------------------------------------------------------------
Marc C. Nicklaus Lab. of Medicinal Chemistry
e-mail: mn1 -8 at 8- helix.nih.gov National Cancer Institute, NIH
Phone: (301) 402-3111 Bldg 37, Rm 5B29
Fax: (301) 496-5839 BETHESDA, MD 20892-4255 USA
WWW: http://www.nci.nih.gov/intra/lmch/MCNBIO.HTM
------------------------------------------------------------------------
========= ========= ========= ======== ****** ******
From: Vikram Varma <varma -8 at 8- nrcbs8.bio.nrc.ca>
Also consider the Pentium Pros - in parallel, you can get
a lot of bang for your buck!!
>>>>>
Vikram Varma
National Research Council of Canada Phone: 613 993 5150
Institute for Biological Sciences FAX: 613 952 9092
Room 3071 100 Sussex Drive, e-mail: varma -8 at 8- nrcbs8.bio.nrc.ca
Ottawa, Ontario K1A 0R6 www: http://nrcbsa.bio.nrc.ca/~varma
<<<<<
========= ========= ========= ======== ****** ******
From: Alexander Hofmann <hofmann -8 at 8- pctc.chemie.uni-erlangen.de>
Some Gaussian-benchmarks
http://www.chem.joensuu.fi/people/juha_muilu/Misc/benchmarks.html
regards
alex
========= ========= ========= ======== ****** ******
From: "M.F.Guest" <M.F.Guest -8 at 8- dl.ac.uk>
Dear Carles,
In response to your mailing, I thought you might find the attached of
value. This is a plain text version of an assessment of a wide variety
of machines in computational chemistry, including the R10000 and a
prototype version of the new HP machine. While it doesnt specifically
include GAUSSIAN and AMBER, I hope that you will find it of use.
Let me know if I can be of any further assistance.
Best regards
Martyn
******************************************************************
* Martyn F. Guest *
* Head, Advanced Research Computing email: m.f.guest -8 at 8- dl.ac.uk *
* CCLRC Daresbury Laboratory FAX : +44 (0)1925 603634 *
* Warrington voice: +44 (0)1925 603247 *
* Cheshire WA4 4AD *
* England, UK *
******************************************************************
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Performance of Various Computers in Computational Chemistry
Martyn F. Guest
September 1996
Abstract
This report compares the performance of a number of different computer
systems using a variety of software from the discipline of
computational chemistry. The software includes matrix operations, a
variety of chemistry kernels from quantum chemistry and molecular
dynamics, and a set of twelve quantum chemistry calculations using the
GAMESS-UK electronic structure program.
The comparison involves approximately fifty computers, ranging from a
Cray YMP-C98 to scientific workstations from IBM, Sun, Hewlett Packard,
Digital and Silicon Graphics, and Pentium Pro-based PCs.
1. Introduction
This report presents a performance evaluation through benchmarking of
a number of different computer systems specifically in the area of
computational chemistry. This work has been ongoing since 1988, when
the intention was to include representative hardware typifying
supercomputers, superminis, workstations and the emerging class of
parallel, novel architecture machines. Throughout the 1980's the cost
effective debate in scientific computing centred on the relative merits
of conventional vector supercomputers [1] and so-called superminis,
machines costing some 10% of supercomputers, and exhibiting some 10% of
supercomputer performance [2]. Over the past 7 years the supermini
category of computational resource has to a large extent disappeared
from both the vocabulary and offerings of most hardware vendors, to be
replaced by high-end workstation servers. We have retained those
machines that originally belonged in the supermini bracket, the Convex
220, the FPS M64/60 [2] and Alliant FX2808, primarily to provide a
historical perspective at the evolution of workstation capabilities,
and how machines that typically fell in the $500K price range are now
outperformed by machines costing some 10% of this figure.
Supercomputers used in this report include the Cray X-MP, Y-MP, YMP/J90
and YMP/C98, and the Convex C3860. A large number of workstations have
been benchmarked, including those from
- IBM, with the Power and Power2 Risc system RS6000-based models,
- Hewlett Packard with the HP/Apollo DN10020 and the PA-RISC based HP
model 9000 series, including the more recent PA8000 CPU. Note that the
latter was housed in a single-processor HP PA/9000-K460, with a 160 Mhz
PA8000 (and not 180 MHz).
- DEC, with the DEC Station 500, the AXP EV4-based 3000 series. and
the EV5-based 600, 800 and 2100 alpha series,
- Silicon Graphics, with the R3000-, R4000-, and R8000-based machines
including the Challenge, Power Challenge, Indy, Indigo and Indigo2,
Crimson, and 4D series, plus machines with the more recent R10000 and
R5000 CPUs.
- Stardent (1520, 3020 and VISTRA 800), and
- Sun 4/370, SPARCstation-2, SPARC-5 and SPARC-10 and the more recent
Hyper- and Ultra-SPARC processors (Ultra-1/140 and 170, and the
Ultra-2/200).
While workstations have, for the majority of scientific applications,
demonstrated their cost-effectiveness against vector supercomputers,
recent developments with PCs suggest that the position of the
workstation as the desktop system of choice is now under threat. We
thus include for the first time one of the most recent offerings from
the PC marketplace, the 200 MHz Pentium Pro.
Parallel, or ``novel architecture'' machines include the iPSC/860 from
Intel, both transputer and i860-based Meiko Computing Surfaces, the
KSR-2 from Kendall Square Research, the Cray T3D and IBM SP2.
Machines featuring in this exercise, together with associated
configurations, are given in the Appendix. We should stress from the
outset that our access to much of the hardware evaluated herein has
been at best short lived, and has often involved the temporary loan or
donation of machines as part of one of the hardware evaluation
exercises run at the Daresbury Laboratory. In many cases these machines
were not optimally configured in terms of either memory, or high speed
disk, and consideration of the results presented here should be viewed
in that light.
Following an introductory evaluation of hardware based on the Whetstone
Benchmark, we present in Sections 2, 3, and 4 results using a variety
of chemistry-oriented software. This 64-bit floating point precision
FORTRAN-based code may be classified into three distinct categories,
each designed to provide a pointer to the relative hardware
capabilities in the discipline of computational chemistry;
1. The first category (the MATRIX Benchmark, section 2) reflects the
dependence of many of the algorithms in the area of electronic
structure calculations on matrix operations, and includes both matrix
multiplication and matrix diagonalization;
2. The second category (the Computational Chemistry Kernels, section 3)
includes four chemistry `kernels', each comprising less than a 1000
lines of FORTRAN code, and intended to be representative of the typical
calculations undertaken in the area of computational chemistry.
Described in more detail in section 3, these kernels include
direct-SCF, molecular dynamics (MD), quantum monte carlo (QMC), and a
Jacobi eigen solver (JACOBI);
3. Finally, we include complete quantum chemistry applications (section
4), using the GAMESS-UK electronic structure program [3]. Twelve
typical applications are included, featuring both conventional Hartree
Fock self-consistent field (SCF) and direct SCF, complete active Space
SCF (CASSCF) and multiconfiguration SCF (MCSCF), configuration
interaction calculations, both direct-CI and conventional table-driven
MRD-CI, Moller Plesset perturbation theory (MP2), and both SCF and MP2
analytic 2nd derivatives.
2. The MATRIX Benchmark
2.1 Whetstone Benchmark
A comparison of the single processor Whetstone performance on a variety
of machines, including vector supercomputers, minisupers,
superworkstations and workstations, together with that obtained on a
single node of various novel architecture machines is given in Table
1. Data provided includes the Mflop performance on a variety of
floating point vector loops (VL=1024), together with the total cpu time
to execute the benchmark, and the MWips performance. The primary aim
of this benchmark is to provide a performance measure of both floating
point (FP) and integer arithmetic; thus while trends in the VL Mflop
ratings are of interest, only a small part of the total CPU time is
actually involved in these operations. The wide variety of standard
functions exercised (abs, sqrt, exp, alog, sin, cos, atan etc.) consume
a far larger fraction of the reported times; note the latter provide a
close mapping onto the measured rate of instruction processing (MWips,
million whetstones instructions per second).
This benchmark was originally designed to monitor the performance of
vector supercomputers, and an examination of Table 1 reveals that both
the Cray Y-MP and YMP-C90 continue to outperform all other machines.
As expected, the Cray YMP-J90 is less impressive, some four times
slower than the C90, and slower than five of the leading workstation
CPUs. The fastest CPU is seen to be the 195 MHz R10000 processor from
SGI, some 1.5 times faster than the 333 MHz EV5 of the alpha 600/5.
Both CPU time and MWips rate suggest that the latter processor is
marginally faster than the recently released PA8000 processor of the HP
PA/9000-K460, and 1.5 times the speed of the 200 MHz Ultra-2/200 from
Sun. The performance of the K460 is far below that expected based on
the published SPEC-fp95 ratings, and it is clear that the compiler,
libraries etc on the machine under test (a single processor 160 Mhz
prototype of the PA8000) were not optimal, with the results of this and
subsequent benchmarks not reflecting the true potential of the PA8000.
SGI's other new processor, the R5000 from mips, is also seen to perform
well, and while about 1.8 times slower than the R10000, is faster than
the Ultra-2 and alpha 266 Mhz EV5. In contrast the R8000-based machines
from SGI appear to under-perform, 4 times slower than the R10000, three
times slower than the AXP-600/5/333, and slower than both the HP
PA/9000-735/125 and IBM Power 2 RS/6000-3CT. The R8000 is seen to
outperform its predecessor, the R4400, by just a factor of 1.7, based
on the CPU time for the complete benchmark. Sun's latest processor,
the 200 MHz Ultra-2/200 is found to be more than twice the speed of the
125 Mhz HyperSPARC, and 8.5 times the 40 MHz SuperSPARC processor in
the Sun SPARC 10/41.
Considering the MPP single node performance of Table 1, the KSR-2
custom processor is seen to be comparable with mid-range workstation
CPUs (IBM RS/6000-360 and DEC AXP/3000-500), while the quality of the
FORTRAN compiler on the Cray T3D is at least in part responsible for
the Whetstone timing of 121 seconds, 1.8 times slower than the DEC
AXP/3000-500 (68.7 secs) which does, of course, house the same CPU.
The IBM-SP2 TN2 processor is by far the dominant CPU, twice the speed
of the KSR-2 and 3.5 times that of the Cray T3D.
Finally, we note that the performance of the Pentium Pro is broadly in
line with expectations. It outperforms the R8000-based SGI and
Power2-based IBM workstations, while slower than the leading CPUs from
SGI, Digital, HP and Sun, by factors of 3.8, 2.5, 2.3 and 1.7
respectively. It is important to realise that this level of performance
is only achieved using "commercial" Fortran compilers (in this case
from Intel). Using public domain f2c and a variety of c-compilers
produced far inferior figures, some 3-4 times slower than those of
Table 1.
2.2 Sparse Matrix Multiply Benchmark
The matrix multiply operation (MMO) is central to the efficient
operation of modern QC codes on vector processors [1], it being
possible both to extract near peak performance for this kernel and to
formulate many QC steps around this operation. A comparison of the
single processor sparse MMO performance on a variety of machines, is
given in Table 2. In this benchmark a series of MMOs (R = A X B)
involving matrices of order 10, 20, 30, ... 150 were performed. Each
MMO was conducted a number of times, this number being inversely
proportional to the order of the matrices, so that the summed CPU times
of Table 2. refer to 150 MMOs of order 10 matrices, 140 of order 20
matrices, and so on up to 10 MMOs for matrices of order 150. Figures
are presented for both `full' (0% sparse) and 50%-sparse B matrices,
with the performance figures referring to code written entirely in
Fortran.
The potential of the PA8000 processor is clearly seen in the figures of
Table 2, with the HP/PA9000-K460 recording the same time as the Cray
Y-MP/C98, and marginally faster than the R10000/195 processor. Both
processors outperform the Cray Y-MP/8128, with the HP/PA9000-K460 twice
as fast as the R8000-based SGI machines and Power2 RS/6000-3CT. The
latter machines exhibit identical performance on this benchmark (with
0% sparsity timings of 1.8 seconds). Both CPUs are marginally faster
than the EV5 600/5/333 and 84000/5/300, 1.3 times faster than the
Ultra-2/200 and 1.4 times faster than the HP PA/9000-735/125. Both
exhibit comparable timings to that on the Y-MP/8128 (1.5 seconds),
achieve some 50% of the YMP-C98 rating, and outperform the Cray YMP-J90
by a factor of 1.4. The performance of the R5000 processor is perhaps
disappointing, four times slower than the R10000. The R1000-based
Power Onyx is seen to outperform the corresponding R8000 machine by a
factor of 1.6, which in turn outperforms the R4400 Challenge L by a
factor of 3.9, and the earlier 50 Mhz R4000 Challenge by a factor of
5.6. Much of the speed of the R8000 may be attributed to the KAP
pre-processor, which enhances FORTRAN performance by a factor of 1.8.
The performance of the Pentium Pro/200 is again seen to be impressive,
recording exactly the same benchmark time as the DEC Alpha 600/5-266
and Sun Ultra-2/200.
Of historical note is the timing of 448.6 seconds recorded on the T800
20MHz Transputer, some 500 times slower than the PA8000 CPU and the
Cray YMP-C98. Considering the MPP nodes, the IBM-SP2 TN2 node
outperforms the Cray AXP node of the T3D by a factor of 4.7, and the
KSR-2 by a factor of 6.6. Note that the T3D node remains 1.5 times
slower than the DEC AXP/3000-500.
One additional feature of the MMO benchmark not apparent in Table 2 is
the significantly enhanced performance found on all machines when
comparing assembly language MMO to Fortran MMO. Historically this has
often involved the user having to code key routines in assembly
language, for few vendors initially provided optimized mathematical
libraries; improvement factors of 3.2 (Cray X-MP), 4.2 (IBM 3090/VF),
3.1 (Convex C-220) and 3.4 (FPS-M64/60) have previously been reported
with assembly language implementations of the sparse MMO routine,
MXMB. Optimized BLAS libraries are now fairly commonplace on the
majority of workstation platforms, the most notable exception, until
recently, being the SPARC offerings from Sun (note that the
Cray-optimised libraries are now available on both Hyper- and
UltraSPARC machines). The impact of optimized library routines
continues to be evident on current workstations, with the 0% sparsity
timings of Table 2 improving by factors of 1.8, 1.9, 2.3, 2.5, 2.7 and
3.1 for the SGI R10000/195, IBM RS/6000-3CT, SGI R8000 Indigo2, HP
PA/9000-735/125, DEC Alpha 600/5/333 and Alpha 8400/5-300 respectively
when using the BLAS dgemm routine. A corresponding factor of 3.8 is
found when using the Kuck Library on the i860, and perhaps somewhat
surprisingly, a factor of 4.5 when using SCILIB on the Cray YMP-C98.
The availability and degree of functionality of library software are,
we believe, important issues when considering cost-effective
performance. This effect is evident from the second benchmark which,
given in Table 3, involves performing a series of similarity transforms
(Q*HQ) using both a scalar and vector algorithm. The scalar code
collapses the matrix transposition and multiplications to yield an
algorithm with fewer FLOPS than the vector code, which adopts a brute
force approach by explicitly performing the transposition and two
matrix multiplications. In the latter case we utilize the BLAS library
routine DGEMM (where available) for performing the requisite MMOs, in
the former case the dot product BLAS routine, DDOT.
Considering the scalar algorithm, the SGI R10000 and HP PA/9000-K460
exhibit the optimum performance, with the R10000 1.1 times the speed of
the Sun Ultra-2/200 and 1.2 times that of the IBM RS/6000-3CT. The
Power2 CPU is marginally faster than the DEC 600/5/333, and 1.2 times
faster than the R8000-based SGI machines and the Sun Ultra-1/170. With
a couple of notable exceptions. this ordering is similar to that found
for the vector case, where the R10000 just outperforms the Sun
Ultra-2/200 and the DEC 600/5/333, which are in turn superior to both
the IBM RS/6000-590 and R8000-based SGI machines. The performance of
the R10000 in the scalar algorithm is first class, faster than the Cray
Y-MP/J90, Cray Y-MP/8128 and Cray YMP C98/4256 by factors of 4.3, 2.3
and 1.2 respectively. The Pentium Pro/200 again fares well, recording
comparable scalar timings to the DEC Alpha 600/5-266, Sun Ultra-1/140,
and DEC Alpha 2100/5-250, although 1.7 times slower than the R10000.
The notably poor performance on the vector algorithm for the HP
PA/9000-K460, SGI Indy 5000 and Pentium Pro/200 may be attributed to
either use of poorly tuned maths libraries (as on the HP and SGI
machine), or to reliance on straight Fortran code (as on the Pentium).
A similar effect was seen on the RS/6000-3CT, where the relatively
unimpressive timings from the vector algorithm were caused by the
un-tuned library DGEMM routine available through -lblas; the
corresponding ESSL routine (-lessl) improves performance by a factor of
1.52 on the IBM SP2. The Cray YMP-C98 at last justifies its
supercomputer tag, outperforming the SGI R10000, R8000 and RS/6000-3CT
by factors of 2.6, 3.2 and 3.6 respectively; the same cannot be said
for the Cray Y-MP/J90, which remains a factor of 1.7 slower than the
R10000. The latter outperforms the R5000 by factors of 3.4 (scalar) and
5.3 (vector).
As noted above all of the more recent CPUs (with the exception of the
PA/9000-K460 and Pentium Pro) exhibit superior performance on the
vector algorithm, with average factors of 2.1 (DEC Alpha 600/5-333),
2.0 (SUN Ultra-1/170), 1.8 (DEC Alpha 8400/300 and HP PA/9000-735), 1.7
(SGI R10000, SUN Ultra-2/200 and SGI R8000), and 1.6 (IBM RS/6000, with
the exception of the 590). Much higher factors are found with the
vector CPUs; the figure of 4.2 on the Cray J90 leads to the Cray being
competitive on the vector algorithm, but slower than the leading 24
workstations on the scalar code. Note again the inadequacies of the
FORTRAN compiler on the Cray T3D; the unexceptional scalar algorithm
timing of 79.3 secs. (to be compared with that of 45.1 secs. on the
DEC AXP/3000-500), improves significantly with the vector algorithm,
where use of the BLAS routines produces timings of 27.5, to be compared
with 23.9 secs on the AXP/3000-500. This effect is also seen in
comparison with the KSR-2 and IBM SP2 timings; the T3D is slower by a
factor of 1.4 on the scalar algorithm, and faster by a factor of 1.3 in
the vector case compared to the KSR-2, while the differential with the
SP2 is reduced from 5.8 in the scalar case to 3.3 for the vector
algorithm.
2.3 Diagonalization Benchmark
Table 4 presents the results of a matrix diagonalization benchmark
intended to supplement the previous analysis conducted by Dunning and
co-workers [4]. We consider a similar benchmark, based on
diagonalizing a series of real symmetric matrices, with rank 10, 20,
30, ... 100, using 64-bit floating point arithmetic. Again the CPU
time was measured for the diagonalization of each size matrix, with the
summed times used as the benchmark execution time. Results are
presented for a range of compiler options available on the depicted
hardware. While the previous analysis was restricted to the EISPACK RS
routine, we consider below the performance of eight diagonalization
routines available in various mathematical libraries and quantum
chemistry codes:
(i) EIGRS, an unoptimized FORTRAN version of the library routine RS
(available in SCILIB on the Cray) from the IMSL library of routines
[5].
(ii) F02ABF from the NAG library [6].
(iii) HQRII [7], as implemented in the semiempirical MOPAC program
[8].
(iv) GIVENS, adapted from the QCPE program exchange (number 62.1).
(v) SDIAG2, as implemented in the MUNICH system of programs.
(vi) JACOBI, from the ATMOL system of programs [9].
(vii) JACO, the diagonalization routine from the direct-SCF program
DISCO [10].
(viii) ERDUW, as taken from the Berkeley System of Quantum Chemistry
codes.
The first four routines are all based on the Householder QR method,
whilst the last four use the Jacobi method. Note that the only
optimization performed involved inserting calls to the BLAS for
two-dimensional rotations (DROT) and vector interchange (DSWAP).
The timings of Table 4 suggest that the SGI R10000/195 is again the
fastest CPU, just ahead of the HP PA/9000-K460 and DEC Alpha 600/5-333,
and 1.5 times faster than the Sun Ultra-2/200 and DEC Alpha
8400/5-300. The R8000-based machines, HP/9000-735/125 and IBM
RS/6000-3CT are seen to exhibit comparable timings (8.0-8.1 secs.), a
factor of 2.5 times slower than the R10000. This benchmark tends to be
dominated by the slowest of the diagonalization routines in use, JACO,
which typically accounts for > 40% of the total CPU time.
Significantly the leading twenty workstations are seen to outperform
the Cray YMP/C98, while the Cray Y-MP/J90 is bettered by the leading 44
workstation CPUs. The Pentium Pro is seen to be five times faster than
the J90, and is only outperformed by the leading six workstation CPUs.
The performance of the Cray T3D node is comparable to the IBM
RS/6000-350, and again significantly slower than the IBM-SP2 TN2 node
(21.7 secs. vs. 9.3 secs.)
2.4 Relative Performance on Matrix Operations
To summarize the performance of the various workstations on the matrix
multiply and diagonalization benchmarks detailed above, we show in
Table 5 the performance of each relative to the SGI R10000/195. It is
clear from these figures that the SGI R10000 and PA8000 processor (in
the HP PA/9000-K460) lie ahead of the competition, with the R10000 1.3
times faster than the DEC Alpha 600/5-333 and 1.4 times faster than the
Sun Ultra-2/200. While the PA8000 is marginally slower than the R10000,
we belief that the arrival of tuned maths libraries on the former will
reverse this order to provide a picture more consistent with the
SPECfp95 ratings.
We see that the relative ordering of processors within a given family
are broadly in line with clock speeds. Considering the EV5-based CPUs,
we find the DEC Alpha 600/5-333, 8400/5-300, 600/5-266 and 2100/5-250
to be slower than the R10000 by factors of 1.25, 1.49, 1.82 and 1.67
respectively. For the Sun Ultra, the Sun Ultra-2/200, Sun Ultra-1/170,
and Sun Ultra-1/140 are slower than the R10000 by factors of 1.43, 1.69
and 2.05 respectively. We also note the following:
- the 266 MHz EV5-based CPU outperforms the corresponding EV4 by a
factor of 1.33;
- the R10000 exhibits a speed up of 1.7 against its predecessor, the
R8000 in the Power Onyx, and a speed up of 3.4 against the R5000 (due
in the main to the poor library performance on the R5000);
- the R8000 exhibits a speed up of 3.3 against its predecessor, the
R4400 in both the Indigo2 and Challenge L, and,
- the position of the Pentium Pro as the 14th fastest CPU of those
considered will undoubtedly improve given the advent of optimised
libraries (not available in this current exercise).
3. Computational Chemistry Kernels
One of the crucial requirements in evaluating the increasingly broad
range of hardware platforms, whether these be parallel machines (true
MIMD message-passing machines, workstation clusters, shared-memory
multiprocessors etc), or simply the lastest workstation, is the
availability of portable benchmarking codes that are representative of
the application area under consideration. In an attempt to provide
such capabilities in computational chemistry, we have described
previously four representative codes, each of which is less than 1000
lines of FORTRAN, and is sufficiently portable that migration to any
hardware platform can typically be achieved in a matter of hours. In
this report we limit our discussion to the performance of these codes
on a variety of single CPUs.
The benchmarking kernels comprise the following programs that are
realistic models of actual chemical applications or algorithms;
1. Self Consistent Field (SCF); This Self Consistent Field (SCF)
electronic structure kernel uses distributed primitive 1s gaussian
functions as a basis (thus emulating use of s,p,... functions) and
computes integrals to essentially full accuracy. It is a direct SCF
code, with an atomic density used for a starting guess. There are two
available problem sizes, corresponding to 60 basis functions Be4 and
240 basis functions (Be16). The timings of Table 6 refer to the
former.
2. Molecular Dynamics (MD); This program bounces a few thousand argon
atoms around in a box with periodic boundary conditions. Pairwise
interactions (Leonard-Jones) are used with a simple integration of the
Newtonian equations of motion.
3. Monte Carlo (MC); This code evaluates the energy of the simplest
explicitly correlated electronic wavefunction for the He atom ground
state using a variational monte-carlo method without importance
sampling.
4. Jacobi iterative linear equation solver (JACOBI)]; Uses a naive
jacobi iterative algorithm to solve a linear equation. All the time is
spent in a large matrix vector product.
Total CPU timings for the SCF, MD and MC benchmarks are presented in
Table 6, together with the Mflop ratings from the JACOBI benchmark.
These results present a somewhat confusing picture, with the processor
ordering very much a function of the particular chemistry kernel under
examination. With the exception of the JACOBI benchmark, the SGI
R10000/195, DEC Alpha 600/5-333 and 8400/5-300, and Sun Ultra-2/200
processors are seen to be the superior CPUs, although the processor
ordering in the SCF, MD and MC benchmarks varies significantly.
Thus for the SCF kernel, the SGI R10000 and Alpha 600/5-333 are 1.5
times the speed of the HP PA/9000-K460, twice the speed of the Sun
Ultra-2/200 and three times that of the POWER2 RS/6000. The
HP/9000-735/125 exhibits good relative performance - it is 1.3-1.4
times faster than the R8000 and POWER2 RS/6000 (TN2 and 3CT). The Sun
SS20/HS21 is also seen to perform well on this kernel, with the same
execution time as the latter CPUs. The performance of the Pentium
Pro/200 matches that of the SGI R8000 and and POWER2 RS/6000.
The SGI R10000 is clearly the optimum processor in the MD benchmark,
followed by the Sun Ultra-2/200 and Alpha 600/5-333. The POWER2 is
seen to perform very poorly in the MD Benchmark where it is apparently
almost 4.5 times as slow as the R10000, with the 3CT 3.8 times slower
than the Sun Ultra-2/200. In fact the whole family of RS6000
processors remains consistently unimpressive on this benchmark.
The relative processor ordering found in the MD and SCF kernels is seen
to be markedly different to that suggested by the JACOBI benchmark.
The Mflop ratings of Table 6 suggest that the Power2 3CT is the optimum
CPU, with the 81 Mflop rating 2.1 times that achieved on the SGI
R10000, 2.5 times that found on the R8000 (32.9 Mflop), and 2.6 times
that recorded on the HP PA/9000-K460 (31.0 Mflop). The Sun Ultra
performance on JABOBI is also impressive, with the Ultra-2/200, and
Ultra-1/170 and Ultra-1/140 all surpassing the MFlop rate on the EV5
and R10000 processors.
In the MC benchmark, the Alpha 600/5-333 and SGI R10000 are the fastest
processors, twice the speed of the RS/6000 POWER2 and Ultra-2/200.
Comparing the R10000/195 and R8000 processors, the R10000 is 3.2 times
faster in the SCF kernel, 2.3 in the MC kernel, but only 1.2 times
faster in JACOBI. The Indy R5000 demonstrates comparable performance to
the R8000 - it is slower than the R10000 in the SCF, MD, MC, and JACOBI
benchmarks by factors of 2.9, 2.1. 1,9 and 3.1 respectively. The
performance of the Pentium Pro/200 is perhaps somewhat less impressive
than that found in the matrix benchmarks. Its overall performance is
similar to that shown by the DEC Alpha 250/4-266 and Indy R5000,
approximately one half the speed of the R10000.
4. The Quantum Chemistry Benchmark
The benchmark described below (and summarized in Table 7) is designed
to highlight the typical range of calculations commonly performed by
the ab initio quantum chemist. It includes 12 calculations carried out
using the GAMESS-UK electronic structure code, and includes the
following functionality;
- Conventional SCF Calculations, on morphine and 2,4,6
tri-nitro-toluene (Calculations 1. and 2. respectively);
- Valence-only ECP calculations, with a geometry optimization of
Na7Mg+ in an ECP-DZ+D(Mg) basis of 70 GTOs (calculation 3).
- A Direct-SCF calculation on cytosine (82 GTOs, 6-31G basis)
Calculation 4);
- Multi-configuration SCF calculations, with a CASSCF geometry
optimization on H2CO (Calculation 5), and a larger CASSCF calculation,
also on H2CO (Calculation 6);
- Configuration interaction calculations, both Direct-CI on the
H2CO/H2+CO transition state (Calculation 7), and conventional table
driven-CI on TiCl4 (Calculation 8);
- Moller Plesset calculations, with a MP2 geometry optimization of
H3SiNCO (Calculation 9);
- Analytic second derivatives, at both the SCF (pyridine) molecule
in a 6-31G basis, Calculation 10) and MP2 level (C4) in a 6-31G* basis,
Calculation 11);
- A Direct-MP2 calculation on pyridine in a DZ + D(N) basis set
(Calculation 12).
4.1 QC Benchmarks - Single Processor Results
The data presented in Table 8 is collected under control of the UNIX
command time where available, and includes CPU time (both user and
system), total elapsed time and Efficiency, measured as CPU versus
elapsed. While our original aim was to base comparisons strictly on
elapsed times, such timings could not be consistently gathered over
the range of machines considered. For example, the range of disk
configurations varies enormously, from primitive SCSI disks to striped
high-speed raid disks, and the loading on the machines varies, from
effectively single-user loading on many of the workstations, to
multi-user environments, such as on the Convex. Thus while reporting
the elapsed times, we use such figures to identify, where appropriate,
the requirement for enhanced disk configurations, rather than as any
definite criticism of the machine in question.
The total user CPU timings of Table 8 suggest that the DEC Alpha
600-5/333 and SGI R10000/195 are the optimum machines, with summed
timings of 23.4 and 23.1 minutes for all 12 calculations. These are
significantly less than those for the leading machines from Sun and IBM
which exhibit summed timings of 31.5 and 40.1 for the Sun Ultra-2/200
and RS/6000-3CT respectively. A somewhat different picture emerges
when considering the system CPU and Elapsed times. With the exception
of the AXP and SGI R10000, all machines exhibit a system CPU time of
the order of 10-15% of the user time; this percentage increases
significantly, particularly on the AXP, to between 20-40% on the
systems of Table 8.
Considering the elapsed times and associated efficiences. the R8000
based machines from SGI appear to be the most balanced. Both the R8000
Power Onyx and Indigo2 exhibit efficiences of > 95%, to be compared
with figures of 76%, 64% and 59% for the RS/6000-3CT, 9000-735/125 and
Alpha 600/5-266. What is noticeable is the significant improvements in
these efficiency ratios on the more recent Sun (Ultra-2/200) and
Digital hardware (both the DEC Alpha 600/5/333 and DEC Alpha
2100/5/250) compared to the figures recorded in the past.
Initial experience with the HP PA/9000-K460 suggested that the compiler
was too unreliable to even attempt to perform the GAMESS-UK Benchmark,
with numerous routines mis-compiling. The poor elapsed time on the Sun
Ultra-1/140 was due in part to the limited memory on the machine (32
MByte).
Overall factors of 10.6 (in total CPU) and 13.1 (in elapsed times) are
found when comparing the ``slowest'' (Sun SPARCstation 10/30) and
``fastest'' (SGI R10000 Power Onyx) machines in Table 8. The
corresponding CPU factors in both the Matrix and Chemistry Kernels are
somewhat higher, 18.9 and 11.2 respectively.
Considering machines from a given vendor, we find that the EV5-based
Alpha 600/5/266 outperforms the corresponding EV4 system, the 250/4/266
by factors of 1.5 (CPU) and 1.4 (elapsed). The R10000 from SG is found
to be approximately twice the speed of the R8000-based machines, and
3.2 times as fast as the Indy R5000. The R8000-based machines in turn
are found to perform some 1.9 times faster than the corresponding Power
Challenge L and R4400 Indigo2. Significantly higher ratios are found in
some of the individual calculations of Table 7, with ratios of 2.5 -
2.8 found in the MCSCF, direct-CI, MP2 and 2nd derivative
calculations. Two of the 12 benchmark calculations impact seriously on
the final ratios, the ECP and MRDCI calculations exhibiting negligible
speedups of 1.13 and 1.18 respectively.
5. Summary
As a summary of this work, we present in Table 9 the relative
performance of 34 of the leading workstations against the DEC Alpha
600/5/333 in terms of quoted SPEC (``Systems Performance Evaluation
Cooperative'') benchmarks, and those from the present Matrix, Chemistry
Kernels and GAMESS-UK benchmarks. Note that this analysis has been
somewhat complicated by the inconsistent approach adopted by some
vendors to providing a smooth transition between the older SPECfp92 and
SPECint92 values, and the more recent SPECfp95 and SPECint95
benchmark. SGI in particular appear to provide only SPECfp95 values
for the R10000 and R5000, and only SPECfp92 values for their older
processors. In the following discussion we have used the SPEC-95
results when available (see Table 9)
The SPEC benchmark suite contains non-tuned application-based code to
measure processor speed for both integer (SPECint) and floating point
(SPECfp) arithmetic. Based on the published SPECfp ratings, and
normalising with respect to the DEC Alpha 600/5/333, we would expect
the PA8000 processor of the HP PA/9000-K460 (124%) to be the fastest
CPU, followed by the DEC Alpha 600/5/333 itself, the SGI R10000/195
(94%) and DEC Alpha 8400/5/300 (94%), the DEC Alpha 600/5/266 CPU
(89%), the 200 MHz Sun Ultra-2/200 (84%), and the IBM Power2
RS/6000-3CT (77%) (where the %-values in parentheses indicate
performance relative to the Alpha 600/5/333). Based on these relative
SPECfp values given in the table, we expect a factor of 12 between the
fastest and slowest processor, the Sun SPARC/10-30. A somewhat
different processor ordering is revealed by the SPECint ratings, with
the processors from SGI (notably the R8000) and IBM (the Power2)
significantly slower than those from DEC, Sun and the HP (PA8000). It
is perhaps worth noting that SGI have not as yet published the
SPECint95 rating for either the R10000 or R5000. We note also that the
Specint95 ratings suggest that Pentium Pro/200 is some 88% of the DEC
Alpha 600/5/333, to be compared with the smaller rating of 51% based on
SPECfp95.
When considering the results, there are several factors we wish to
consider based on the present evaluation exercise:
1. Do the SPECfp values provide a reliable metric for evaluating the
capabilities of hardware in computational chemistry? If this so, we
would expect to find a close mapping of the ratios for the various
chemistry benchmarks onto the SPECfp ratios (note that the CPU values
for the GAMESS-UK benchmark are used, since SPEC does not adequately
incorporate I/O in its evaluation);
2. Does any particular CPU consistently ``underperform'' based on the
SPECfp criteria? - this would manifest itself as the ratios from the
chemistry benchmarks falling below the SPECfp ratios;
3. Do the ``simple'' Matrix and Chemistry Kernel benchmarks lead to the
same conclusions as the GAMESS-UK benchmarks?
In terms of relative speed, we find that all the chemistry benchmarks
are broadly in line with the SPECfp predictions, the only notable
exceptions being summarised below:
- The poor performance of the HP PA/9000-K460 - while the matrix
kernels are in line with the SPECfp ratio, the chemistry kernels would
appear to be running at around half the expected speed.
- The impressive performance of the SGI R10000 on all the benchmarks,
with figures of 130%, 104% and 116% on the matrix, kernels and
GAMESS-UK benchmarks, as against the value of 94% based on the SPECfp95
figures. Indeed, the R10000 would certainly appear to be the optimum
CPU based on the benchmarks conducted in this report. The R5000
appears, however, to follow closely the SPECfp95 ratings.
- There is some evidence from Table 9 that neither the DEC EV5-based
Alpha 600/266 or the 8400/300 are performing as well as the SPEC
ratings might suggest, with all benchmark ratios some 10% lower than
expected based on SPECfp alone. In contrast the DEC Alpha 2100/5/250 is
performing consistently above its SPEC rating.
- The UltraSPARC systems from Sun appear to be performing exactly in
line with the SPECfp95 ratings.
- The R8000-based systems perform well on the matrix benchmarks, but
are closer to the SPECfp ratings on the kernels and GAMESS-UK.
- All IBM Power1 RS/6000 CPUs exhibit enhanced performance on the
chemistry codes relative to the SPEC rankings e.g.. the SPECfp ratio of
the RS/6000-370 to DEC 600/5 is 22%, to be compared with the Matrix,
Kernels and GAMESS-UK ratios of 30%, 37% and 36% respectively.
- A similar effect is shown with the HP PA/9000-735/125, where the
SPECfp ratio of 35% increases to 54%, 40% and 52% for the chemistry
benchmarks.
- The potential of the Pentium Pro is evident, with the performance in
the Matrix benchmarks exceeding the SPECfp95 figures.
References
[1] M.F. Guest and S. Wilson, Daresbury Laboratory Preprint,
DL/SCI/P290T; Supercomputers in Chemistry, ed. P.Lykos and I.Shavitt,
A.C.S. Symposium series 173 (1981) 1; V.R. Saunders and M.F. Guest,
Comp.Phys.Comm., 26 (1982) 389: M.F. Guest, in ``Supercomputer
Simulations in Chemistry'', Ed. M. Dupuis, Lecture Notes in Chemistry,
44, Springer Verlag (1986) 98.
[2] M.F. Guest, R.J. Harrison, J.H. van Lenthe and L.C.H. van Corler,
``Computational Chemistry on the FPS-X64 Scientific Computers:
Experience on single- and multi-processor systems'', Theoret. Chim.
Acta 71 (1987) 117.
[3] GAMESS-UK is a package of ab initio programs written by M.F. Guest,
J.H. van Lenthe, J. Kendrick, K. Schoeffel and P. Sherwood, with
contributions from R.D. Amos, R.J. Buenker, M. Dupuis, N.C. Handy, I.H.
Hillier, P.J. Knowles, V. Bonacic-Koutecky, W. von Niessen, R.J.
Harrison, A.P. Rendell, V.R. Saunders, and A.J. Stone. The package is
derived from the original GAMESS code due to M. Dupuis, D. Spangler
and J. Wendoloski, NRCC Software Catalog, Vol. 1, Program No. QG01
(GAMESS), 1980.
[4] R. Shepard, R.A. Bair, R.A. Eades, A.F. Wagner, M.J. Davis, C.B.
Harding and T.H. Dunning Jr., Int. J. Quant. Chem. , (1983) 17; R.A.
Bair and T.H. Dunning Jr., J.Comp.Chem., 5 (1984) 44; R. Bair,
``FPS-164 Matrix Multiplication Subroutine Guide'', Argonne National
Laboratory, (1984); T.H. Dunning Jr. and R.A. Bair, ``Advanced Theories
and Computational Approaches to the Electronic Structure of
Molecules'', NATO ASI Series, D. Reidel, 1984, p1.
[5] IMSL Program Library, 1978.
[6] NAG Fortran Library, Numerical Algorithms Group Ltd, 1984.
[7] Y. Beppu, Computers and Chemistry, 6 (1982).
[8] J.P. Stewart, ``MOPAC - A General Molecular Orbital Package'', QCPE
455, 1984.
[9] V.R. Saunders and M.F. Guest, ``ATMOL3 Part 9'', RL-76-106, 1976;
M.F. Guest and V.R. Saunders, Mol.Phys., 28 (1974) 819; D. Moncrieff
and V.R. Saunders, ``ATMOL-Introduction Notes'', UMRCC, May, 1986;
Cyber-205 Note Number 32, UMRCC, September, 1985.
[10] J. Almlof et al, J.Comp.Chem., 3 (1982) 385.
Performance of Various Computers in Computational Chemistry
Table 1. Vector Whetstone Benchmark. Total CPU times (seconds), Mflop
ratings (see text) and MWIP ratings
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Machine Mflop Ratings (VL=1024) Total CPU MWIPS
N2 N3 N8 (seconds)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
iPSC/860 (RX) 4.6 4.4 5.3 282.7 15.0
Cray T3D (AXP) 9.8 8.4 7.3 120.5 33.5
KSR-2 32.8 31.3 29.7 68.6 59.6
IBM SP2 (TN2 node) 49.2 38.2 49.3 35.0 116.5
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Sun 4/370 2.0 1.8 1.0 642.8 6.3
SOLBOURNE S4000 1.4 1.3 0.9 629.3 6.4
Sun SPARCstation 2/GS 4.3 3.2 1.8 366.7 11.0
DEC S5000/120 4.5 4.0 2.7 351.3 11.5
IBM RS/6000-320 9.4 2.4 1.3 305.3 13.3
SGI 4D/220 5.6 4.4 2.9 289.5 14.0
Stardent 1520 6.3 5.9 6.0 283.2 14.5
DEC S5000/200 6.4 5.4 3.4 274.7 14.8
SGI R3000 Indigo 6.8 3.7 3.7 233.8 17.3
SGI 4D/320 7.6 5.9 3.6 227.0 18.0
Sun SPARC/5-85 3.3 3.1 4.3 205.5 19.7
Stardent VISTRA-800 4.7 4.4 5.3 196.5 19.0
SGI 4D/420 9.4 6.9 4.4 188.0 21.6
Sun SPARC/10-30 7.0 3.4 4.0 177.2 21.0
Sun SPARC/10-41 9.4 8.1 5.3 145.0 26.5
SGI IRIS Crimson 21.8 14.3 6.1 121.7 33.7
IBM RS/6000-540 15.1 7.7 5.7 120.3 33.8
HP PA/9000-720 11.6 11.1 9.3 119.3 34.0
IBM RS/6000-340 17.9 4.0 6.1 115.9 34.8
Sun SPARCserver 1000 12.5 10.9 6.5 111.6 33.8
SGI R4000 Indigo 21.8 13.6 6.4 105.8 38.1
IBM RS/6000-530H 16.4 10.4 6.8 102.0 39.9
RS/6000 PowerPC-250 15.1 8.4 5.4 95.0 38.7
IBM PowerPC-25T 15.1 8.4 6.4 93.2 41.1
HP PA/9000-730 14.0 14.2 11.5 93.4 43.4
SGI Challenge L/100 21.8 13.6 9.1 90.8 43.7
IBM RS/6000-550 21.8 10.8 7.9 86.8 46.6
IBM RS/6000-350 19.7 8.0 7.7 84.8 46.9
DEC AXP/3000-300 19.4 15.7 10.3 79.1 50.9
IBM RS/6000-360 24.6 9.0 9.2 70.7 56.1
DEC AXP/3000-500 20.8 16.9 10.7 68.7 58.7
DEC AXP/3000-600 26.2 19.8 12.7 64.4 62.8
HP PA/9000-750 15.1 27.5 24.3 64.1 63.6
SGI R4400 Indigo^2 32.7 19.9 13.1 62.3 63.8
SGI Challenge L/150 32.8 20.5 14.2 62.9 63.3
IBM RS/6000-370 32.8 12.0 11.5 56.5 70.2
HP PA/9000-715/100 32.8 29.9 15.8 54.5 73.3
Sun SPARCstation-20/HS21 30.0 25.5 18.2 49.8 81.6
HP PA/9000-J200 39.3 32.8 33.9 41.0 99.2
HP PA/9000-755 39.3 44.4 30.9 38.9 104.7
DEC AXP/3000-700 34.1 25.6 23.6 36.5 111.1
RS/6000 PowerPC-43P 21.9 9.8 24.2 36.4 92.2
SGI R8000 Indigo^2 27.0 26.2 109.2 36.4 112.4
SGI R8000 PowerOnyx 27.1 26.2 109.6 36.2 112.8
IBM RS/6000-590 32.8 31.3 44.2 34.8 117.3
IBM RS/6000-3CT 39.3 37.2 46.8 34.4 117.6
IBM RS/6000-3BT 39.2 37.2 47.2 34.2 118.9
Pentium Pro/200 41.8 35.2 21.7 32.3 123.7
HP PA/9000-735/125 39.3 57.3 38.9 30.8 132.0
DEC Alpha 250/4-266 36.6 34.8 72.5 26.9 150.5
Sun Ultra-1/140 19.9 19.3 170.9 26.3 155.1
Sun Ultra-2/200 29.8 27.0 239.6 18.7 218.2
DEC Alpha 2100/5-250 143.9 39.4 107.4 17.0 238.1
DEC Alpha 600/5-266 83.9 83.9 126.3 15.8 256.1
DEC Alpha 8400/5-300 94.4 97.9 133.5 14.2 286.7
HP PA/9000-K460 196.6 196.6 190.5 13.8 298.3
DEC Alpha 600/5-333 201.4 52.4 138.7 12.8 311.0
SGI PowerOnyx 10000/195 118.6 76.0 139.2 8.5 476.8
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Convex C-220 21.4 17.7 15.7 59.6 70.0
Convex C-3860 54.0 44.3 37.8 23.0 185.2
CRAY YMP/J90 115.9 117.9 145.9 16.6 260.2
CRAY Y-MP/8128 190.0 193.5 297.2 7.9 552.0
CRAY YMP-C98/4256 542.1 539.3 668.7 3.9 1114.0
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Table 2. Sparse MMO Benchmark: Total CPU times (seconds) for a series
of sparse MMOs (R = A X B, see text) implemented in Fortran.
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Machine Sparsity in B-Matrix
0% 50%
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
iPSC/860 RX-node 21.2 11.0
KSR-2 11.3 5.9
Cray T3D AXP-node 8.5 4.5
IBM SP2 TN2-node 1.8 1.1
Solbourne S4000 74.4 38.4
Sun 4/370 57.4 30.6
DEC S5000/120 50.7 25.9
HP/Apollo DN10020 44.2 22.8
SGI 4D/220 39.4 20.1
DEC S5000/200 33.1 17.1
SGI R3000 Indigo 31.0 14.3
SGI 4D/320 27.9 14.4
Sun SPARCstation 2/GS 27.5 14.4
Stardent 1520 26.6 17.0
SGI 4D/420 23.7 12.1
Stardent VISTRA-800 20.4 10.6
Sun SPARCstation 10/30 16.1 7.9
Sun SPARCstation 5/85 14.9 7.8
Stardent 3020 14.7 9.1
IBM RS/6000-320 14.6 7.1
SGI R4000 Indigo 12.6 6.4
SPARCstation 10/41 11.8 6.1
SGI IRIS Crimson 11.0 5.6
SGI Challenge L/100 10.0 4.9
IBM RS/6000-530 9.3 4.9
Sun SPARCserver 1000 9.3 4.7
IBM PowerPC-25T 9.2 4.3
RS/6000 PowerPC-250 9.1 4.4
IBM RS/6000-340 8.9 4.1
HP PA/9000-720 8.5 4.8
IBM RS/6000-530H 7.8 4.1
IBM RS/6000-350 7.6 3.6
SGI R4400 Indigo^2 7.1 3.4
SGI Challenge L/150 7.1 3.4
HP PA/9000-730 6.7 3.5
IBM RS/6000-360 6.4 3.1
DEC AXP/3000-300 6.1 3.3
DEC AXP/3000-500 5.7 2.9
IBM RS/6000-550 5.6 3.1
HP PA/9000-750 5.1 2.7
IBM RS/6000-370 5.1 2.4
DEC AXP/3000-600 5.0 2.6
RS/6000 PowerPC-43P 4.3 2.0
HP PA/9000-715/80 3.7 2.0
DEC AXP/3000-700 3.7 1.8
Sun SPARCstation 20/HS21 3.5 1.8
Sun Ultra-1/140 3.4 1.7
DEC Alpha 250/4/266 3.2 1.6
HP PA/9000-755 3.1 1.7
HP PA/9000-715/100 3.0 1.6
Sun Ultra-1/170 2.8 1.4
DEC Alpha 2100/5/250 2.6 1.3
HP PA/9000-735/125 2.5 1.3
Sun Ultra-2/200 2.3 1.2
DEC Alpha 600/5-266 2.3 1.2
Pentium Pro/200 2.3 1.2
HP PA/9000-J200 2.1 1.1
DEC Alpha 8400/5-300 2.1 1.1
IBM RS/6000-3BT 2.1 1.1
IBM RS/6000-590 1.9 1.0
DEC Alpha 600/5-333 1.9 1.0
SGI R8000 Indigo^2 1.8 1.1
SGI R8000 Power Onyx 1.8 1.1
IBM RS/6000-3CT 1.8 1.0
SGI PowerOnyx 10000/195 1.1 0.6
HP PA/9000-K460 0.9 0.5
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Alliant FX2808 (1CE) 17.2 9.3
FPS-M64/60 17.1 8.9
Convex C-220 10.6 6.5
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
IBM 3090-600-E/VF 11.3 6.0
Convex C-3860 5.2 3.3
Cray X-MP/416 2.8 1.8
Cray YMP/J90 2.6 1.8
Cray Y-MP/8128 1.5 0.9
CRAY YMP C98/4256 0.9 0.6
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Table 3. Sparse MMO Benchmark: Total CPU times (seconds) for a series
of Similarity Transformations (H=Q*HQ, see text) using both Scalar and
Vector Algorithms.
Machine Algorithm
Scalar Vector
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
T800-20 5176.0 4439.0
iPSC/2 SX-node 3809.6 4456.1
Meiko MK086 node 240.1 262.2
iPSC/860 RX-node 118.8 60.1
KSR-2 55.3 34.8
Cray T3D AXP-node 79.3 27.5
IBM SP2 TN2-node 13.8 8.5
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
SOLBOURNE S4000 711.9 750.2
Dec S5000/120 710.2 724.6
Sun 4/370 615.9 609.6
Dec S5000/200 368.3 445.1
SGI 4D/220 344.8 440.3
Sun SPARCstation 2/GS 333.3 394.0
SGI R3000 Indigo 285.0 356.4
SGI 4D/320 245.5 329.5
HP/Apollo DN10020 273.8 245.9
Stardent 1520 236.9 252.8
SGI 4D/420 200.1 273.8
Stardent VISTRA-800 177.0 209.4
Stardent 3020 160.4 144.2
Sun SPARCstation 10/30 141.5 162.3
Sun SPARCstation 10/41 111.3 119.3
Sun SPARCstation 5/85 105.9 151.6
Sun SPARCserver 1000 88.8 92.6
SGI Indigo R4000 98.2 78.7
SGI IRIS Crimson 94.2 78.7
IBM RS/6000-320 139.0 73.3
IBM PowerPC-25T 88.2 68.9
RS/6000 PowerPC-250 92.5 68.7
HP PA/9000-720 124.7 66.5
SGI Challenge L/100 74.9 58.2
HP PA/9000-730 89.9 48.4
IBM RS/6000-530 96.3 46.9
IBM RS/6000-340 70.4 43.4
HP PA/9000-750 58.6 41.5
SGI Indigo^2 R4400 52.3 39.8
SGI Challenge L/150 56.5 39.0
IBM RS/6000-540 66.9 38.4
IBM PowerPC-43P 44.3 36.4
Sun SPARCstation 20/HS21 42.2 35.1
IBM RS/6000-350 55.8 34.8
HP PA/9000-715/80 52.7 34.4
DEC AXP/3000-300 61.0 34.0
IBM RS/6000-530H 58.8 34.0
SGI Indy R5000 34.7 32.1
HP PA/9000-715/100 47.0 29.5
IBM RS/6000-360 48.0 29.3
IBM RS/6000-550 48.3 27.7
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
DEC AXP/3000-500 45.1 23.9
HP PA/9000-735 34.8 20.0
IBM RS/6000-370 37.6 23.3
DEC AXP/3000-600 38.5 20.2
HP PA/9000-755 35.6 20.1
Pentium Pro/200 17.1 23.2
HP PA/9000-J200 33.5 16.2
IBM RS/6000-3BT 30.8 16.2
HP PA/9000-735/125 29.2 16.2
DEC AXP/3000-700 28.3 16.2
DEC Alpha 250/4-266 24.2 13.1
HP PA/9000-K460 10.9 12.0
DEC Alpha 600/5-266 17.7 10.2
IBM RS/6000-3CT 12.6 9.7
Sun Ultra-1/140 17.1 9.5
SGI R8000 Indigo^2 15.2 8.7
SGI R8000 Power Onyx 15.1 8.7
DEC Alpha 2100/5-250 17.7 8.2
Sun Ultra-1/170 15.5 7.9
DEC Alpha 8400/5-300 13.8 7.9
Sun Ultra-2/200 11.5 6.8
DEC Alpha 600/5-333 13.2 6.2
SGI PowerOnyx 10000/195 10.2 6.1
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Alliant FX2800 (1CE) 243.1 86.3
FPS-M64/60 141.2 50.8
CONVEX C-220 128.6 49.0
CONVEX C-3860 55.1 20.1
Cray YMP/J90 44.1 10.6
Cray Y-MP/8128 23.2 5.7
Cray YMP C98/4256 12.6 2.6
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Table 4. Matrix Diagonalization Benchmark. Total CPU times (seconds)
for a series of Matrix Diagonalizations (see text) using eight
Different Routines.
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Machine CPU Time Compiler
(seconds) Options
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
T800-20 1100.5
iPSC/2 (SX) 1059.0 -OLM
MK086 i860 66.8 -OLM
iPSC/860 (RX) 62.3 -O3
KSR-2 27.3 -O2
Cray T3D (AXP) 21.7 -O
IBM SP2 (TN2) 9.3 -O3
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Stardent 1520 255.0 -O2
Sun 4/370 219.9 -O
SOLBOURNE S4000 169.0 -O3
DEC S5000/120 145.8 -O2
Sun SPARCstation 2/GS 103.3 -O
SGI 4D/220 97.4 -O2
Apollo DN10020 96.2 -O
DEC S5000/200 90.6 -O2
SGI R3000 Indigo 78.8 -O2
Stardent 3020 68.1 -O2
SGI 4D/320 67.2 -O2
Stardent VISTRA 59.2 -O3
SGI 4D/420 58.1 -O2
IBM RS/6000-320 46.4 -O
Sun SPARCstation 10/30 42.2 -O -dalign
Sun SPARCstation 10/41 35.9 -O -dalign
IBM RS/6000-530 35.6 -O
Sun SPARCstation 5/85 32.4 -O -dalign
HP PA/9000-720 31.9 -O
IBM RS/6000-540 30.8 -O
SGI IRIS Crimson 27.6 -sopt
IBM RS/6000-340 27.6 -O
IBM RS/6000-530H 26.0 -O
SGI R4000 Indigo 26.0 -O -mips2
RS6000 PowerPC-250 25.1 -O
Sun SPARCserver 1000 24.7 -O -dalign
HP PA/9000-730 24.3 -O
IBM PowerPC-25T 24.0 -O3
IBM RS/6000-550 22.2 -O
IBM RS/6000-350 21.3 -O
HP PA/9000-750 20.8 +O3
SGI Challenge L/100 20.4 -O -mips2
IBM RS/6000-360 17.7 -O
HP PA/9000-715/80 15.9 +O3
SGI Challenge L/150 15.6 -O -mips2
DEC AXP/3000-300 15.5 -O
SGI R4400 Indigo^2 15.2 -O -mips2
DEC AXP/3000-500 14.9 -O
IBM RS/6000-370 14.2 -O3
HP PA/9000-715-100 12.8 +O3
Sun SS20/HS21 12.4 -O -dalign
Sun SPARCstation 20/HS21 12.4 -O -dalign
DEC AXP/3000-600 12.1 -O
IBM PowerPC-43P 11.2 -O3 qarch=ppc
HP PA/9000-735 10.0 +O3
HP PA/9000-755 10.2 +O3
HP PA/9000-J200 9.6 +O4
IBM RS/6000-3BT 9.4 -O3 qarch=pwr2
IBM RS/6000-590 9.0 -O3 qarch=pwr2
DEC AXP/3000-700 8.7 -O
IBM RS/6000-3CT 8.1 -O3 qarch=pwr2
HP PA/9000-735/125 8.1 +O3
SGI R8000 Indigo^2 8.0 -O3 -mips4
SGI R8000 Power Onyx 8.0 -O3 -mips4
SGI Indy R5000 7.7 -O3 -mips4
DEC Alpha 250/4-266 7.3 -fast
Sun Ultra-1/140 6.2 -fast -O4
DEC Alpha 600/5-266 5.6 -fast
Sun Ultra-1/170 5.4 -fast -O4
Pentium Pro/200 5.3 -G6 -O2
DEC Alpha 2100/5-250 5.2 -fast
DEC Alpha 8400/5-300 4.6 -fast
Sun Ultra-2/200 4.5 -fast -O4
DEC Alpha 600/5-333 3.9 -fast
HP PA/9000-K460 3.5 +Oaggressive
SGI PowerOnyx 10000/195 3.1 -O3 -mips4 etc
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Convex C-220 86.5 -O2
Alliant FX2808 99.5 -Og
FPS-M64/60 80.7 OPT3
Convex C-3860 36.0 -O2
Cray X-MP4/16 21.3
Cray YMP/J90 26.7
Cray Y-MP4/64 16.0
Cray YMP C98/4256 9.9
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Table 5. The Matrix Benchmark: Performance relative to the SGI
Power Onyx R10000/195
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Machine MMO MMO Diagonal Total
(FORTRAN) (Q*HQ) -ization
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
SGI R10000/195 100% 100% 100% 100%
HP PA/9000-K460 123% 57% 90% 90%
DEC Alpha 600/5-333 61% 99% 80% 80%
Sun Ultra-2/200 50% 90% 70% 70%
DEC Alpha 8400/5 300 55% 78% 68% 67%
DEC Alpha 2100/5 250 44% 75% 60% 60%
Sun Ultra-1/170 41% 78% 58% 59%
SGI R8000 Power Onyx 63% 71% 39% 58%
SGI R8000 Indigo^2 63% 71% 39% 58%
IBM SP2 (TN2 node) 63% 72% 35% 57%
DEC Alpha 600/5-266 50% 60% 56% 55%
IBM RS/6000-3CT 63% 64% 38% 55%
Sun Ultra-1/140 33% 64% 50% 49%
Pentium Pro/200 50% 36% 59% 48%
HP/9000-J200 55% 38% 32% 42%
IBM RS/6000-3BT 54% 38% 33% 42%
DEC Alpha 250/4-266 36% 47% 43% 42%
HP PA/9000-735/125 46% 38% 39% 41%
Cray YMP/J90 44% 58% 12% 38%
DEC AXP/3000-700 31% 38% 36% 35%
HP PA/9000-755 37% 31% 31% 33%
SGI Indy R5000 25% 22% 41% 29%
HP PA/9000-715/100 38% 21% 34% 28%
DEC AXP/3000-600 23% 30% 26% 26%
Sun SparcStation-20/HS21 33% 17% 25% 25%
IBM PowerPC-43P 27% 17% 28% 24%
IBM RS/6000-370 22% 26% 22% 24%
HP PA/9000-715/80 31% 18% 20% 23%
DEC AXP/3000-500 20% 26% 21% 22%
IBM RS/6000-360 20% 21% 18% 20%
DEC AXP/3000-300 19% 18% 20% 19%
IBM RS/6000-550 20% 21% 14% 19%
HP PA/9000-750 22% 15% 15% 18%
SGI R4400 Indigo^2 16% 15% 21% 17%
SGI Challenge L/150 16% 15% 20% 17%
Cray T3D AXP 13% 22% 14% 16%
IBM RS/6000-350 15% 18% 15% 16%
IBM RS/6000-340 13% 14% 11% 13%
SGI Challenge L/100 11% 11% 15% 12%
IBM PowerPC-25T 12% 9% 13% 12%
Sun SPARCserver-1000 12% 7% 13% 11%
HP PA/9000-720 13% 9% 10% 11%
SGI R4000 Indigo 9% 8% 12% 10%
Sun SparcStation-10/41 10% 6% 9% 8%
Sun SparcStation-5/85 8% 6% 10% 8%
IBM RS/6000-320 8% 8% 7% 8%
Sun SparcStation-10/30 7% 4% 7% 6%
Stardent VISTRA 6% 3% 5% 5%
SGI 4D/420 5% 3% 5% 4%
SGI 4D/320 4% 3% 5% 4%
SGI R3000 Indigo 4% 2% 4% 3%
Sun SPARCstation-2 4% 2% 3% 3%
DEC S5000/200 3% 2% 3% 3%
Stardent 1520 4% 3% 1% 3%
Apollo DN10020 3% 2% 3% 3%
SGI 4D/220 3% 2% 3% 3%
DEC S5000/120 2% 1% 2% 2%
Solbourne S4000 2% 1% 2% 1%
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Table 6. Computational Chemistry Kernels
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Machine Total CPU time (secs) Mflop
=====================
SCF MD MC Jacobi
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Sun SPARCstation 10/30 758 2214 299 6.9
Sun SPARCstation 10/41 665 1788 266 6.6
Sun SPARCstation 5/85 1036 1082 195 6.7
Sun SPARCserver 1000 441 783 148 6.7
RS6000 Power-PC 250 603 1322 149 15.6
IBM RS/6000-530H 649 1594 176 26.7
IBM RS/6000-340 692 1608 187 21.2
IBM Power-PC 25T 409 1322 145 15.6
SGI Challenge L/100 400 794 130 9.5
SGI R4000 Indigo 375 809 112 9.5
DEC AXP/3000-300 427 716 98 12.0
IBM RS/6000-350 523 1288 141 27.4
HP PA/9000-750 237 554 163 11.5
%IBM RS/6000-550 563 1266 149 39.9
IBM RS/6000-360 433 1071 117 35.1
HP PA/9000-715/80 220 427 133 17.1
SGI Challenge L/150 233 517 78 13.3
SGI R4400 Indigo^2 234 523 78 13.7
DEC AXP/3000-600 314 541 79 21.6
HP 9000/715-100 177 354 107 17.3
IBM Power-PC 43P 135 470 56 14.8
IBM RS/6000-370 344 854 93 44.2
HP PA/9000-J200 121 292 92 12.4
Sun SPARCstation 20/HS21 132 319 84 16.3
HP PA/9000-755 131 284 102 18.6
DEC AXP/3000-700 146 411 54 21.7
HP PA/9000-735/125 103 236 91 15.0
SGI Indy R5000 124 219 42 12.5
SGI R8000 Indigo^2 139 218 49 23.7
Pentium Pro/200 142 321 38 28.4
SGI Power Onyx R8000 138 218 49 32.9
DEC Alpha 250/4/266 79 272 37 20.9
IBM/RS6000-3BT 140 389 43 44.8
Sun Ultra-1/140 110 186 61 45.2
IBM RS/6000-590 146 453 43 86.0
Sun Ultra-1/170 98 146 58 47.3
DEC Alpha 2100/5/250 59 187 28 27.3
IBM SP2 TN2-node 138 383 43 68.6
HP PA/9000-K460 68 138 31 31.0
DEC Alpha 600/5/266 57 210 31 40.9
IBM RS/6000-3CT 138 462 42 81.1
Sun Ultra-2/200 82 121 49 56.5
DEC Alpha 8400/5/300 50 183 27 43.9
DEC Alpha 600/5/333 46 137 21 44.9
SGI PowerOnyx 10000 43 106 22 39.1
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
CONVEX C-220 1015 1017 329 33.1
CONVEX C-3860 453 342 141 69.9
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Table 7. The GAMESS-UK Single Processor Benchmark
Number Module Basis (GTOs) Details Molecular
Species
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1. SCF STO-3G (124) Morphine
2. SCF 6-31G (154) C6H3(NO2)3
3. ECP Geometry ECPDZ (70) Na7Mg+
Optimization
4. Direct-SCF 6-31G (82) Cytosine
5. CASSCF TZVP (52) 480 csf H2CO
Geometry Opt.
6. MCSCF (5s3p2d/3s1p/f(O) 5608 H2CO
(74)
7. Direct-CI (5s3p2d/3s1p) 3M/167194 H2CO/H2+CO TS
(64)
8. Table-CI (26M/6R) ECP (59) 2301815/4097 TiCl4
9. MP2 Geometry 6-31G* (70) H3SiNCO
Optimization
10. SCF Second 6-31G (64) C5H5N
Derivatives
11. MP2 Second 6-31G* (60) C4
Derivatives
12. Direct-MP2 DZ+D(N) (76) C5H5N
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Table 8. The GAMESS-UK Benchmark: Total CPU time (user and system)
Elapsed time (minutes) and Efficiency (%) for Calculations 1 - 12
(see text)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Machine CPU Time Elapsed Efficiency
User System Time (%)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
DEC Alpha 600/5/333 23.4 7.6 34.4 91%
SGI PowerOnyx 10000 23.1 3.6 36.7 73%
Sun Ultra-2/200 31.5 3.0 38.1 91%
DEC Alpha 2100/5/250 30.2 8.2 43.6 88%
SGI R8000 Power Onyx 48.1 5.6 55.2 97%
Sun Ultra-1/170 40.5 4.9 56.3 81%
IBM RS/6000-3CT 40.1 4.4 58.3 76%
SGI Indigo^2 R8000 50.8 7.2 60.6 96%
IBM RS/6000-590 46.6 4.7 74.8 68%
DEC Alpha 600/5/266 30.6 6.4 62.2 59%
DEC Alpha 250/4/266 46.0 9.7 85.7 65%
DEC Alpha 8400/5/300 26.9 10.4 92.8* 40%
HP PA/9000-735/125 51.6 8.4 93.3 64%
HP PA/9000-755 64.9 8.8 98.2 75%
SGI Indy R5000 72.9 11.3 101.0 83%
DEC AXP/3000-700 53.4 11.5 105.4 62%
DEC AXP/3000-600 72.1 14.1 107.8 80%
IBM RS/6000-370 78.7 8.4 108.8 80%
SGI R4400 Challenge L/150 93.3 10.3 110.6 94%
SGI R4400 Indigo^2 94.3 10.0 120.6 86%
HP 9000/715-100 81.4 14.3 124.7 77%
Sun Ultra-1/140 45.3 5.7 139.8* 36%
HP PA/9000-715/80 93.7 15.9 140.1 78%
DEC AXP/3000/500 93.0 40.1 148.6 90%
HP PA/9000-750 111.1 17.4 140.2 92%
IBM RS/6000-550 117.6 11.2 146.4 88%
SGI Challenge L/100 140.8 17.1 178.0 89%
DEC AXP/3000-300 129.8 39.6 184.0 92%
Sun SS-20/HS21 82.5 16.2 202.7 49%
SGI R4000 Indigo 153.2 17.7 206.0 83%
Sun SPARCserver 1000 175.1 50.8 445.6 51%
Sun SPARCstation 10/41 201.4 21.0 260.4 85%
Sun SPARCstation 5/85 240.5 25.8 299.0 89%
Sun SPARCstation 10/30 251.4 31.7 306.3 92%
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
CONVEX C-3860 124.3 3.5 210.9 61%
Cray Y-MP/8128 56.8 6.0 188.4 33%
Cray YMP-C98/4256 37.9 2.0 41.2 97%
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
* see footnotes to Table 9
Table 9. The Chemistry Benchmark: Performance relative to the
DEC Alpha 600/5/333 (see text)
Machine SPECfp SPECint Matrix CC GAMESS-UK
Kernels CPU Wall
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
DEC Alpha 600/5/333 100 100 100 100 100 100
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
HP PA/9000-K460 124 113 124 76
SGI R10000/195 94 130 104 116 93
DEC Alpha 8400/5/300 94 81 85 85 83 37 *
DEC Alpha 600/5/266 89 86 71 76 84 55
Sun Ultra-2/200 84 83 87 86 90 90
IBM RS/6000-3CT 77 37 72 73 70 59
Sun Ultra-1/170 69 60 73 70 68 61
DEC Alpha 2100/5/250 64 65 74 71 81 79
Sun Ultra-1/140 60 50 61 63 61 25 +
SGI Power Onyx $ 57 26 75 53 58 62
Pentium Pro/200 51 88 64 48
SGI Indigo^2 R8000 $ 49 27 75 48 54 57
DEC Alpha 250/4/266 48 56 53 53 56 40
DEC 3000/700 AXP $ 43 40 45 38 48 32
SGI Indy R5000 36 39 44 37 34
HP 9000/735-125 35 43 54 40 52 37
HP 9000/755 $ 31 26 43 36 42 35
DEC 3000/600 AXP $ 30 28 33 29 36 17
Sun SS-20/HS21 $ 28 32 34 35 31 17
DEC 3000/500 AXP 28 23 28 19 23 23
HP 9000/715-100 26 31 38 31 32 27
HP 9000/715-80 $ 23 23 31 27 28 24
IBM RS/6000-370 $ 22 17 30 37 36 31
DEC 3000/300 AXP $ 17 16 25 20 18 19
SGI Challenge L/150 $ 16 22 22 26 30 31
SGI Indigo^2 R4400 $ 16 22 23 26 30 28
SPARC SS-1000 $ 15 14 14 14 8
HP 9000/750 $ 14 12 24 21 24 24
IBM RS/6000-550 $ 13 9 24 31 24 23
SGI Challenge L/100 $ 12 15 16 17 20 19
SGI Indigo R4000 $ 11 14 13 17 18 17
Sun SS-10/41 10 12 11 9 14 13
Sun SS-5/85 $ 10 16 10 11 12 11
Sun SS-10/30 $ 10 11 8 9 11 11
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
$ - SPECfp and SPECint rations based on SPEC92 values.
* - multi-user environment on DEC Alpha 8400/5/300.
+ - small memory configuration (32 MByte) on Sun Ultra-1/140.
Table 10. APPENDIX I: Machine Configurations under Evaluation
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Machine Configuration Location
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Workstations
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Solbourne S4000 DL (loan)
Sun 4/370 DL
Sun SPARCstation 2/GS SPARC/40 Mhz DL (loan)
Sun SPARCstation 5/85 MicroSPARC II/85 MHz DL (loan)
Sun SPARCstation 10/30 SuperSPARC/36 MHz DL (loan)
Sun SPARCstation 10/41 SuperSPARC/40 MHz PNL
Sun SPARCserver 1000 SuperSPARC/50 MHz DL (loan)
Sun SPARCstation 20/HS21 HyperSPARC/125 MHz DL(loan)
Sun Ultra-1 Model 140 UltraSPARC-1/143 MHz DL (loan)
Sun Ultra-1 Model 170 UltraSPARC-1/167 MHz DL
Sun Ultra-2 Model 200 UltraSPARC-2/200 MHz DL (loan)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
HP/Apollo DN10020 PRISM DL
HP PA/9000-720 PA7000/50 MHz DL
HP PA/9000-730 PA7000/66 MHz MCC
HP PA/9000-750 PA7000/66 MHz DL
HP PA/9000-715/100 PA7100LC/100 MHz DL (loan)
HP PA/9000-715/80 PA7100LC/80 MHz DL (loan)
HP PA/9000-755 PA7100/99 MHz DL
HP PA/9000-735/125 PA7150/125 MHz PNL
HP PA/9000-J200 PA7200/100 MHz DL (loan)
HP PA/9000-K460 PA8000/160 MHz DL (loan)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
DEC S5000/120 R3000A/R3010A 20 Mhz DL (loan)
DEC S5000/200 R3000A/R3010A 25 MHz DL (loan)
DEC AXP/3000-300 AXP A21064/150 MHz PNL
DEC AXP/3000-500 AXP A21064/150 MHz DL (loan)
DEC AXP/3000-600 AXP A21064/175 MHz PNL
DEC AXP/3000-700 AXP A21064A/225 MHz DL (loan)
DEC Alpha 250/4/266 AXP A21064A/266 MHz DL (loan)
DEC Alpha 600/5/266 AXP A21164/266 MHz DL (loan)
DEC Alpha 2100/5/250 AXP A21164/250 MHz DL (loan)
DEC Alpha 8400/5/300 AXP A21164/300 MHz RAL
DEC Alpha 600/5/333 AXP A21164/333 MHz DL (loan)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
SGI 4D/220 GTX DL
SGI 4D/320 R3000A/R3010A 33 MHz DL (loan)
SGI 4D/420 DL (loan)
SGI IRIS Crimson R4000/R4010 100 Mhz DL (loan)
SGI R3000 Indigo R3000A/R3010A 33 MHz DL (loan)
SGI R4000 Indigo R4000/R4010 100 MHz DL (loan)
SGI Challenge L/100 R4400/R4010 100 MHz Utrecht
SGI R4400 Indigo^2 R4400/R4010 150 MHz DL (loan)
SGI Challenge L/150 R4400/R4010 150 MHz Southampton
SGI R8000 Indigo^2 R8000/R8010 75 MHz DL (loan)
SGI R8000 Power Onyx R8000/R8010 75 MHz DL (loan)
SGI Indy R5000 R5000/R5000 180 Mhz DL (loan)
SGI PowerOnyx 10000 R10000/R10010 195 MHz DL (loan)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Stardent 1520 DL
Stardent 3020 DL (loan)
Stardent VISTRA-800 DL (loan)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
IBM Power1 RS/6000-320 DL (loan)
IBM Power1 RS/6000-530 DL (loan)
IBM Power1 RS/6000-530H RS6000/33 MHz DL
IBM Power1 RS/6000-540 Perugia
IBM PowerPC-25T MPC601 66 MHz DL
IBM PowerPC-43P MPC604 100 MHz DL
IBM Power1 RS/6000-340 RS6000/33 MHz DL (loan)
IBM Power1 RS/6000-350 RS6000/41.6 MHz DL (loan)
IBM Power1 RS/6000-550 RS6000/41.6 MHz Perugia
IBM Power1 RS/6000-360 RS6000/50 MHz DL (loan)
IBM Power1 RS/6000-370 RS6000/62.5 MHz DL
IBM Power2 RS/6000-590 RS6000/66 MHz IBM
IBM Power2 RS/6000-3BT RS6000/67 MHz DL (loan)
IBM Power2 RS/6000-3CT RS6000/72 MHz DL (loan)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
PCs
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Netpower PC Pentium Pro/200Mhz DL (loan)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
SuperMinis
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Alliant FX2808 DL
FPS-M64/60 DL
Convex C-220 DL
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
SuperComputers
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Convex C-3860 ULCC
IBM 3090-600-E VF RAL
Cray YMP J90/10 EPCC
Cray Y-MP/8128 SARA
Cray YMP C98/4256 SARA
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Novel architecture Node
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Meiko Computing Surface T800-20 DL
iPSC/2 SX-node DL
iPSC/2 VX-node DL
iPSC/860 RX-i860 node (40 MHz) DL
KSR-2 KSR-2 node (80 MHz) PNL
Cray T3D AXP node (150 MHz) Edinburgh
IBM SP2 TN2 node (67 MHz) DL
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
========= ========= ========= ======== ****** ******
_______________________________________________________________________
Carles Colominas
Dept. of Org. Chem. E-mail ccolo -8 at 8- mompou.iqs.url.es
Institut Quimic de Sarria. TEL: (34-3)-203.89.00
Universitat Ramon Llull. FAX: (34-3)-205.62.66
Via Augusta 390.
08017-Barcelona. CATALONIA.
_______________________________________________________________________