A Study of Communication Mechanisms for System Area Network on PC Cluster

• Main Authors: Shuen-Tai Wang, 王順泰
• Other Authors: JEN-SHIUH(J.S) LIU
• Format: Others
• Language: zh-TW
• Published: 2001
• Online Access: http://ndltd.ncl.edu.tw/handle/45041279711344774686

碩士 === 逢甲大學 === 資訊工程學系 === 89 === The rapidly reduction of the "Killer Micro" speed made possible the PC cluster to outperform supercomputers. In fact, the supercomputer repackaged these microprocessors in a dense configuration to construct very large parallel computing systems. PC clusters are becoming one of the most important means and clustering technology has been applied for massive high-performance computing by taking advantages of high speed interconnection networks. The goal of this thesis is to study the communication mechanisms and implement some communications software for System Area Network (SAN) on PC cluster. The high data transfer rate is the key factor for high-performance parallel computing in SAN. Unfortunately, this performance has not been realized by current general networks suchas Fast Ethernet. For this reason, we employed Myrinet as our interconnection network, because Myrinet demonstrates the highest performance per unit cost of which we are aware. Also, Myrinet is ideal for cluster-computing applications. There are many communications software running on different platforms now. We studied GAM (Generic Active Message) and Fast Sockets, two commonly used communications software, and ported to our cluster system. Besides, we also built IP/GM, MPICH/GM, and two DSM programming libraries. Review our system, our cluster can provide users a high-performance parallel programming environment. For evaluating the network performance of SAN, we used a NetPIPE microbanchmark for evaluating different communication layers, and modified to work under our cluster. On the other hand, we also used HPL (High Performance Linpack) and ScaLAPACK banchmarks to obtain the overall performance. According to the experimental results, we found that PC clusters will become the building blocks of choice for high-performance parallel computing systems. Cluster of PCs that interconnect by high-speed network will promise to be the supercomputer of the future because of their high-performance and low-cost.