Ab Initio Quantum Chemistry on the IBM pSeries 690

In this study, we compare the performance of the POWER3 processor and the new IBM eServer pSeries 690. The pSeries 690 can scale up to 32-way POWER4 processor at 1.3 GHz and 1.1 GHz. To perform this comparison we used the Gaussian98 Revision A.11 series of electronic structure programs. It is an integrated system to model molecular systems under a variety of conditions, carrying out its calculations from basic laws of quantum mechanics. We analyze and compare the performance of methods such as Hartree-Fock and density functional theory (DFT), including first and second derivatives. In addition, we also look at excitation calculations via CIS. This set of benchmarks indicate that the Turbo 1.3 GHz is approximately 15% faster than the Turbo 1.1 GHz. Scalability between these two machines is similar. Introduction The combined effort between software and hardware is what makes it possible to carry out calculations that a few years ago were unthinkable. An RHF/6-31G** single point energy calculation on triamino-trinitro-benzene (TATB) using Gaussian88 on an IBM RS/6000 model 550 used to take 4.5 hours [1]. The same calculation using Gaussian92 took about 1 hour [1]. Today this calculation takes less than 2 minutes using Gaussian98 on the latest IBM POWER3 systems. In addition to faster processors, chemists have long realized that parallel computing requires scalable software [2] and hardware [3]. Scalable hardware can be classified in terms of how memory is organized [4,5]: distributed-memory, in which each node has its own local memory (e.g., IBM SP nodes ); and shared-addressable memory among all the processors (shared memory). The IBM pSeries 690 is a major step that IBM has taken toward providing customers with shared memory machines. The next generation of the IBM POWER series (pSeries 690) was introduced in October 2001. This new machine scales up to 32-way POWER4 1.3 GHz or 1.1 GHz SMP and up to 256 GB of memory. The IBM pSeries 690 uses multi-chip modules (MCM). The MCM is equivalent to a mainframe’s central processing module. This approach optimizes chip-to-chip communications, boosting performance of the overall system. The module-to-module interconnect runs at half of the processor’s frequency [6]. Gaussian98 can take advantage of both memory approaches. Previous studies have reported the performance of Gaussian running on different architectures [3,7,8,9]; however, it is beyond the scope of this work to review Gaussian implementation on distributed architectures. In the next section we present the design features of the systems tested on this study. We also provide a section that describes how Gaussian has been parallelized on a shared memory architecture. In the last four sections we describe the benchmarks and their performance. We conclude this work with a section on resource management and a summary. IBM ~ Performance Technical Report Ab Initio Quantum Chemistry on the IBM pSeries 690 Page 2