Research On Operating System Kernel For Scientific Calculation Based Applications

The performance of a supercomputer is mainly decided by hardware, system software, algorithms and application software. Appropriate optimizations and customizations on system software for each category of applications can increase application performance on supercomputers.Scientific calculation is an important usage of supercomputers. Applications based on scientific calculation usually have huge working set, heavy system load, and therefore need a lot of time to finish. The influences of the operating system on performance of scientific calculation are discussed in this dissertation, including not only the influences of OS'memory management and process scheduling strategies, but also the impact of noises generated in OS kernel. And some brief analysis and experiments are carried out upon a few possible optimizing methods on the operating system for scientific calculation, based on above discussion result.A super page allocating mechanics for user programs is implemented in this dissertation, as a supplement of the memory management subsystem in Linux OS. It creates super page frames with legacy 4KB size page frames, and directly assigns them to user programs through reallocated system calls, which are done by a kernel module supported in Linux. This mechanics is not transparent to application programmers, as they can decide by themselves whether super pages should be used in their programs, or adopt both super pages and legacy ones simultaneously. Then a series of performance evaluations are performed on this super page based memory management optimization. The results show that our optimization has a positive impact on the performance of inter-node communication and the performance of address translation in local virtual memory.This dissertation analyzes the inter-node communication pattern of High Performance Linpack (HPL) benchmark program on Myrinet system. The time consumption of communication and the packets distribution in length are observed and examined, as well as the impact of problem size N and data block size NB on the communication pattern. The characteristics of HPL benchmark program's inter-node communication pattern in clusters are then determined. With the understanding of its communication pattern and our super page optimization applied, a HPL performance benchmark is performed on the Dawning 4000A supercomputer. An over 4 percent peak performance improvement is observed comparing to the result obtained without super page optimization, under 1,024 CPUs'scale.The main contribution of this dissertation are: (1) Dawning 4000A supercomputer's two places increment in the TOP500 List of June 2004 (with 10th place); (2) about 17.5 percent increment on the bandwidth of Myrinet when using 8MB packets.