GAMESS/PC GAMESS benchmarks on Different CPUs: A Graphical Analysis.

Valentin Ananikov
GAMESS/PC GAMESS benchmarks on Different CPUs: A Graphical Analysis

Introduction

In this article we present a graphical analysis of the benchmarking data compiled earlier in order to find the main factor(s) the total system performance depends on. The initial data is available as a separate www article - INTEL CPUs Benchmarks under Linux: GAMESS and PC GAMESS programs. In addition, the article contains the necessary information concerning software and hardware details, therefore these questions are not further discussed here.

Job type dependence

The following four job types were executed in all tested systems: PCM Energy, Geometry Optimization, Hessian, and MP2 Geometry Optimization [benchmark data]. The subsequent analysis has shown the total independence of the performance from the job type (fig.1), since the Relative Performance increases in same way for all four calculation types in each Test Number. Very similar picture were obtained in the case of PC GAMESS program [PC GAMESS job type dependence].
The result seems to be obvious inside the same CPU family, however it might differ if the comparison will be made with another vendor products.

Figure 1. Relative Performance on different job types obtained with GAMESS ab-initio program.

*Relative Performance*	The performance of the slowest system (i.e. Pentium 100) were assumed as 1.0 for each job and the rest timings were recalculated on it's bases. For example, point 3 on the Relative Performance axis means that the system is 3 times faster than Pentium 100.
*Test Number*	1 - Intel Pentium 100 MHz 2 - Intel Pentium 166 MHz 3 - Intel Pentium PRO 180MHz 4 - Celeron 300A 300MHz 5 - Intel Pentium II 350MHz 6 - Intel Pentium II 400MHz 7 - Celeron 300A 450MHz
*Colors*	red - PCM Energy green - Geometry Optimization blue - Hessian cyan - MP2 Geometry Optimization

CPU Frequency

    Obviously, that CPU Frequency is one of the main limiting factors of the total system performance, however in the case of Floating Point operations the performance scales linearly (fig.2). Almost perfect line can be obtained if we exclude a bit old Pentium 100 and Pentium 166 from the linear regression fig.3, all the rest CPUs are based on the same chip core.
    PentiumPro, PentiumII and Celeron CPUs have L2 cache of 256Kb, 512Kb and 128Kb correspondingly. The L2 cache on the Celeron chip works at the full CPU frequency, while in the Pentium II chips at the half speed. Nevertheless, the performance is not effected with the difference in the cache type and size. This does not mean that L2 cache has no influence at all, it just indicates that extending a 128Kb Celeron cache (at a full CPU speed) to a 512Kb Pentium II cache (but at a half speed) very likely will not give valuable performance improvement.
    Again, PC GAMESS calculations were resulted in the rather similar pictures [PC GAMESS Frequency dependence].

Figure 2. Averaged Relative Performance for all four job types vs. Absolute CPU Frequency (see also captions to fig.1).

Figure 3. Averaged Relative Performance for all four job types vs. Absolute CPU Frequency (see also captions to fig.1). First two points (100 and 166 MHz) were excluded from linear regression analysis.

Links

GAMESS homepage	http://www.msg.ameslab.gov/GAMESS/GAMESS.html
PC GAMESS homepage	http://classic.chem.msu.su/gran/gamess/
FPU Performance	http://www4.tomshardware.com/releases/99q1/990223/cpu-news-14.html
CPU Site	http://cpusite.examedia.nl/
CPU Guide	http://www4.tomshardware.com/guides/cpu/index.html
INTEL Pentium III site	http://www.intel.com/PentiumIII/
INTEL Pentium II site	http://www.intel.com/PentiumII/
INTEL Celeron site	http://www.intel.com/Celeron/

April 26, 1999