Multiprocessor machine

The above is an example of how direct algorithms may be formulated for methods involving electron correlation. It illustrates that it is not as straightforward to apply direct methods at the correlated level as at the SCF level. However, the steady increase in CPU performance, and especially the evolution of multiprocessor machines, favours direct (and semi-direct where some intermediate results are stored on disk) algorithms. Recently direct methods have also been implemented at the coupled cluster level. 

Any machine with virtual memory suffers performance degradation when page swaps are too frequent. Some pipelined machines like the TI-ASC (3) or the CDC Star-100 (4) have rather long setup times for their arithmetic pipes. Multiprocessor machines like the Illiac IV (5) are next to useless if the programmer pays no attention to the architecture. These features all directly impact the user they have not been effectively hidden by software at any level. Improvement of this situation could result if compilers took on the burden of optimizing code so as to promote efficient hardware utilization. 

Use of a Multiprocessor Machine with a Known Interval of Uncertainty... 

The BzzMinimizationMono class does not use parallel computing however, the BzzMinimizationMonoMP class does and is valid for all cases from the previous class a multiprocessor machine, an adequate compiler, and opmMP directives are required to make proper use of shared memory. 

Chapter 1 discusses methods for handling the function root-finding problem. Algorithms are proposed in the renewed forms to exploit the multiprocessor machines. The use of parallel computing is also common to the chapters that follow. 

The first stage of our prediction algorithm applies the MCM-based approach described above to each of the nine secondary structure predictions for each target. Simulations are usually carried out on two to four nodes of a multiprocessor machine and take between 12 and 24 hours depending on protein size. To extract the structurally unique predictions, we apply the clustering algorithm discussed above. Table VIII shows the results of this procedure for the three targets discussed in more detail below. We list results for every secondary structure prediction (unless predictions consist only of loop or cod, in which case we did not believe it worthwhile to carry out the simulation). 

Currently, work is underway to split up the subsystem level computations across a large number nodes on multiprocessor machines. Each node is assigned the tasks of matrix diag-onalization and density matrix construction for one or more subsystems. Sample timings on the protein crambin for a number of parallel platforms are given in Table 6. From these pilot studies it is clear that some parallelization can be achieved, but that load balancing and parallelization of other parts of the code is required prior to developing an efficient parallel implementation of DivCon. 

Finally, although we have demonstrated that D C calculations can be extremely efficient, we concede that parallelization of the method is necessary in order to make multi-point calculations routine for systems containing thousands of atoms. To this end, we have begun work on a straightforward parallelization procedure to split up the subsystem level computations across nodes on multiprocessor machines. In the very near future, it should be possible to carry out energy minimizations on small proteins in a reasonable amount of time with or without solvent present. 

MIMD Multicomputers. Probably the most widely available parallel computers are the shared-memory multiprocessor MIMD machines. Examples include the multiprocessor vector supercomputers, IBM mainframes, VAX minicomputers. Convex and AUiant rninisupercomputers, and SiUcon... 

The Sony SDP-1000 [Sony, 1989] was an interesting multiprocessor designed around a serial crossbar interconnect. The controlling machine itself can be divided into three sections ... 

Furthermore, hardware like multiprocessor workstations, which provide near-supercomputer performance within the UPSM programming model, are becoming available from several vendors (see chapter appendix). These machines are capable of exploiting the shared-memory parallelism that is already represented in code libraries such as LAPACK. Another important positive sign is that issues of scalable library construction have become more visible—for example, as an IEEE-sponsored workshop. " Such efforts, combined with the availability of software like ScaLAPACK as seed code, may well serve to crystallize the development of common data layout and program structure conventions. 

S. Hiranandani, K. Kennedy, and C. Tseng, in Compilers and Runtime Software for Scalable Multiprocessors, J. Saltz and P. Mehrotra, Eds., Elsevier, Amsterdam, 1991. Compiler Support for Machine-Independent Parallel Programming in FORTRAN D. 

It is well-known that due to administration overhead multiprocessor systems never reach their theoretical performance bound of N times compared to a monoprocessor machine. This can easily be understood by the following consideration. 

FIGURE 1 The performance of sequential and parallel computers between the years 1940 and 2000. The small crosses (x) give the performance of multiprocessor versions of the singleprocessor machines below them marked by a filled circle ( ). The number next to a cross gives the number of processors. For massively parallel machines (o) the number of processors Is given after the forward slash ( ) following the name. 

A switch to connect the nodes. The Butterfly series of shared memory parallel computers manufactured by BBN made use of a type of switch shown in Fig. 12 which connected each processor to every memory unit. Switch-type connections are currently used in symmetric multiprocessors (SMPs). An SMP is a parallel computer in which each processor has equal access to aU I/O devices (including memory). SMPs form the basis of IBM s ASCI White computer, which is currently the fastest in the world. The ASCI White machine is made up of 512 SMP nodes, each with 16 processors, for a system total of 8192 processors. 

---