Harnessing the power of the new SMP cluster architecture

摘要

In 1993, members of our team collaborated with Silicon Graphics to perform the211u001efirst full-scale demonstration of the computational power of the SMP cluster 211u001esupercomputer architecture. That demonstration involved the simulation of 211u001ehomogeneous, compressible turbulence on a uniform grid of a billion cells, using 211u001eour PPM gas dynamics code. This computation was embarrassingly parallel, the 211u001eideal test case, and it achieved only 4.9 Gflop/s performance, slightly over half 211u001ethat achievable by this application on the most expensive supercomputers of that 211u001eday. After four to five solid days of computation, when the prototype machine had 211u001eto be dismantled, the simulation was only about 20% completed. Nevertheless, this 211u001ecomputation gave us important new insights into compressible turbulence and also 211u001einto a powerful new mode of cost-effective, commercially sustainable 211u001esupercomputing (S). In the intervening 6 years, the SMP cluster architecture has 211u001ebecome a fundamental strategy for several large supercomputer centers in the US, 211u001eincluding the DOE's ASCI centers at Los Alamos National Laboratory and at the 211u001eLawrence Livermore National Laboratory and the NSF's center NCSA at the 211u001eUniversity of Illinois. This SMP cluster architecture now underlies product 211u001eofferings at the high-end of performance from SGI, IBM, and HP, among others. 211u001eNevertheless, despite many successes, it is our opinion that the computational 211u001escience community is only now beginning to exploit the full promise of these new 211u001ecomputing platforms. In this paper, we will briefly discuss two key architectural 211u001eissues, vector computing and the flat multiprocessor architecture, which continue 211u001eto drive spirited discussions among computational scientists, and then we will 211u001edescribe the hierarchical shared memory programming paradigm that we feel is best 211u001esuited to the creative use of SMP cluster systems. Finally, we will give examples 211u001eof recent large-scale simulations carried out by our team on these kinds of 211u001esystems and point toward the still more challenging work which we foresee in the 211u001enear future.