As of today, the Windows operating system (OS), when compared with Linux or Unix, has not attracted enough attention regarding building high-performance Beowulf clusters. One important reason is that the performance and reliability of Windows clusters is not clear.
This paper compares the performance of one Linux and two Windows implementations, using the Numerical Aerospace Simulation Parallel Benchmark (NPB) and Parkbench suites. The author claims that “under Windows it is possible to obtain better peak performance[,] but also the computation is extremely unstable and the probability of crash is very high.” A notable fact is that it took the authors more than six months to complete the test using Windows, but less than a week using Linux. It is not clear, however, whether this is due to OS problems, or the message passing interface (MPI) implementation.
I should point out that the data provided in the paper is very limited, in terms of justifying the above conclusion. The weakest aspect of the paper is that there is no deep analysis of the reasons causing the performance difference, so it is not a good paper on a performance study. Another drawback is that the scale of the platform is too small: it has only eight central processing units (CPUs).