Which Operating System to Choose?

The core of GAMESS has been created when the PC industry was in the stage of infancy, and the performance of then existing PCs was barely enough to carry out the Hückel calculation for benzene. Nowadays, the situation has completely changed, and cheap desktop systems are routinely used for serious computations and visualization. The forthcoming discussion only concerns this "cheap iron". One or two x86 processors and IDE hard drives are assumed. A question remains which operating system to boot in order to squeeze the last ounce of performance from your $1000 toy. University folk, including the creators of GAMESS, would probably advise you to stick to Linux. Granovsky suggests Windows NT. I used to run GAMESS under Windows 95 (OSR 2) without any problems, but also changed to NT4 after purchasing a 22GB IBM hard drive. There are a few tips based on my own experience:

Windows 95. If you have got a single-processor single-disk machine with less than 4 GB disk storage and want to have flawless back-compatibility with older DOS and Windows 3.1 applications, you will like Windows 95. Go for OSR 2 version to support large drives (FAT32 file system). Notice however that under FAT32 the maximum file size is limited to 4 GB. Using a robust third-party memory manager, such as QEMM97, and confining the disk cache size to 25% available physical memory (add 'MaxFileCache=' entry in [vcache] section in the system.ini) helps to reduce paging activity, making Windows 95 as fast as NT4. Although I'm an advocate of Windows 95, basically due to its superior software support, beware of possible problems with Windows 98, where large MP2 jobs may cause system instability - apparently, heavy math wasn't the scenario for this gamer-oriented system.

Windows NT4. It is the only choice for power users who own double-processor systems, have more than 1GB RAM, want to maximize the speed of I/O operations by using stripe sets, or need to run heavy computations creating files larger than 4 GB. After all, Windows NT is a more powerful and robust desktop operating system than Windows 95/98. Using third-party disk optimizers, such as Defrag offered by O&O Software or Diskeeper offered by Executive Software, and bus-master IDE drivers introduced in the Service Pack 4 and later, significantly improves execution time of GAMESS jobs. In particular, enabling bus-mastering on my IBM hard-drive speeds up conventional SCF computations by a factor of more than three, whereas such thing as which form of DMA is used - DMA mode 2, UDMA mode 2 (a.k.a. UDMA-33), or UDMA mode 4 (a.k.a. UDMA-66) - has only a marginal effect on performance. Unfortunately, despite Microsoft's advocacy - and the fact that Microsoft Office and some graphical programs indeed run faster under NT4, Windows NT4 itself does not give any noticeable performance boost to GAMESS, MOPAC, and other math-intensive programs mostly relying on the floating-point unit, for the speed of the latter is independent of the operating system. For properly configured Windows 95 and NT4 installations, the execution time does not differ by more than 1%. However, the file caching algorithm used by Windows NT4 proves to be much more effective as compared to that in Windows 95, so there is no need to alter NT's defaults. For example, the CleverCache utility marketed by O&O Software and claimed to double the system performance does nothing but slows down the execution of GAMESS; the degradation of performance is especially noticeable if you're running several tasks at once. Tinkering with NT's memory management will degrade your system as well (for example, when I tried to specify explicitly the amount of L2 cache for Pentium II (512KB), Celeron A (128KB), and Pentium III (256KB), systems by altering the SecondLevelDataCache entry in the Registry, the system performance didn't change at all or even appeared to degrade slightly. An explanation of this I've found quite recently in the Microsoft knowledge base. The thing is that the value of SecondLevelDataCache parameter only affects usage of the so-called direct mapped L2 caches. Thus, a slight performance improvement (two percent) could be detected in older computers (486 and earlier) with memory of 64 MB and more due to the fact that physical pages are distributed better in the address space, reducing L2-cache collisions. Neither Pentium II nor later processors have direct mapped L2 caches. So, don't waste your time. A little knowledge is a dangerous thing, indeed.

Linux. You will like it because it's free, and GAMESS needs no modification to run. Linux is also the most suitable solution for large computer clusters. However, for stand-alone x86 based workstations, I would not pin myself down to Linux. It is not as friendly as Microsoft's products, and even worse, it lacks quality software. Although it is commonly asserted that Linux is faster for numerical tasks, in practice you may often see just opposite: thus, GAMESS compiled with Microsoft PowerStation FORTRAN 4.0 or Digital Visual FORTRAN 5.0 under Windows 95/NT4 runs faster than its f2c/gcc sibling under Linux. For properly configured systems (bus-mastering and memory management are the crucial issues) the difference can be more than a factor of two.