Cray machines

The electronic computer is undeniably an essential component in the tool bag of the modern theoretical chemist and, with ever improved accessibilty to more powerful computing, it is tempting to feel a sense of euphoria about current computer capabilities. At Los Alamos, we have a truly impressive resource of large computers — at present we have a choice of two 1-million word CRAY machines and four CDC 7600 machines. However, even with this powerful a computing environment, the main objective of this talk is to ask whether or not this current euphoria is really justified. Have we yet reached the stage where we can, with current computers, address problems which are traditionally thought of as chemistry The answer is, in many cases, no. 

Our multipole code D-PMTA, the Distributed Parallel Multipole Tree Algorithm, is a message passing code which runs both on workstation clusters and on tightly coupled machines such as the Cray T3D/T3E [11]. Figure 3 shows the parallel performance of D-PMTA on a moderately large simulation on the Cray T3E the scalability is not affected by adding the macroscopic option. 

The Fourier sum, involving the three dimensional FFT, does not currently run efficiently on more than perhaps eight processors in a network-of-workstations environment. On a more tightly coupled machine such as the Cray T3D/T3E, we obtain reasonable efficiency on 16 processors, as shown in Fig. 5. Our initial production implementation was targeted for a small workstation cluster, so we only parallelized the real-space part, relegating the Fourier component to serial evaluation on the master processor. By Amdahl s principle, the 16% of the work attributable to the serially computed Fourier sum limits our potential speedup on 8 processors to 6.25, a number we are able to approach quite closely. 

The most commercially successful of these systems has been the Convex series of computers. Ironically, these are traditional vector machines, with one to four processors and shared memory. Their Craylike characteristics were always a strong selling point. Interestingly, SCS, which marketed a minisupercomputer that was fully binary compatible with Cray, went out of business. Marketing appears to have played as much a role here as the inherent merits of the underlying architecture. 

Mathematical models require computation to secure concrete predictions. Successes in relatively simple cases spurs interest in more complex situations. Somewhat specialized computer hardware and software have emerged in response to these demands. Examples are the high-end processors with vector architecture, such as the Cray series, the CDC Cyber 205, and the recently announced IBM 3090 with vector attachment. When a computation can effectively utilize vector architecture, such machines will out-perform even the most powerful conventional scalar machine by a substantial margin. Such performance has given rise to the term supercomputer. ... 

To the disappointment of Lilly s guest relations department, Lilly s Cray-2 was later replaced with a Cray J90, a mundane-looking machine. But the J90 was more economical, especially because it was leased. The supercomputers were almost always busy with molecular dynamics and quantum mechanical calculations [100]. Of the personnel at the company, the computational chemists were the main beneficiaries of supercomputing. 

Support by the Division of Astronomical Sciences of the National Science Foundation is gratefully acknowledged. The model calculations were performed on the Cray YMP8 computer of the Ohio Supercomputer Center we are grateful for the award of time on this machine. 

The CRAY-1 vector processing computer at the Science Research Council s (S.E.R.C) Daresbury Laboratory, is at the centre of a network providing large scale computational facilities for Universities in the United Kingdom. This is the only supercomputer available at present to Quantum Chemists in the U.K., and this article will therefore be restricted to experience gained on the CRAY-1, although this experience will undoubtedly be relevant to future applications on machines such as the ICL Distributed Array Processor (DAP) (see reference (2) for a detailed description) and the CDC Cyber 203/205. 

The Daresbury Laboratory CRAY-1 computer is accessed by means of an IBM 370/165 which is linked to computers at the S.E.R.C s Rutherford Laboratory, C.E.R.N. and workstations in many Universities. The S.E.R.C. network in fact incorporates links to 10 mainframe and 76 minicomputers and to 44 different sites. The CRAY-1 was installed at Daresbury for an initial period of two years, extendable for a third year. The S.E.R.C. buys an average of eight hours per day from CRAY Research Inc. Ltd., and the possibility exists that the machine will be upgraded to a CRAY-1 Model S/500. Proposals are also under consideration for the installation of supercomputers at the two largest University regional computer centres - London and Manchester - and at the S.E.R.C s Rutherford Laboratory where the existing twin IBM 360/195 machines are scheduled for replacement in 1982/3. Again these machines would be accessible via workstations in a number of University departments around the U.K. 

Several general lessons may be learned from this study. The kernel of the integrals code actually operates at approximately 30 times faster than the CDC 7600/ so that the non-kernel code, which takes about 10% of the computer time on a conventional machine occupies 60% of the time in our present CRAY-1 program. The implication is that work to vectorise the non-kernel components of the calculation might now be profitable. 

It is worth pointing out that the vectorised code is written in standard FORTRAN. The CRAY FORTRAN compiler simply recognises the vectorisable loops and translates these into hardware vector orders. An inspection of the machine code thus generated revealed that very little was to be gained by hand coding the kernels into Assembly language. 

As a result of the transfer of the GF Method of sensitivity analysis to the vector machine, an improvement of more than a factor of 100 in running time has been achieved, with an associated cost effectiveness of about 60 from Scheme I on the Honeywell to Scheme II on the Cray. This has been accomplished not only by virtue of the use of a higher speed machine and the vector processor, but also by making the proper choice among alternative algorithms and paying close attention to coding details. 

---