Preemptive Process Migration System With Socket Migration Supporting

With the improvement of performance and the decrease of the price of the commodity part of the computer systems, the cluster systems which have better scalability and higher performance price ratio become more and more important in the high performance computing area. There are plentiful computing resources in high performance computing clusters, therefore managing and utilizing such resources efficiently is a big problem. Dynamic load balancing system that distributing system load among the cluster to get better resource utilization and to improve the overall performance is an important tool to solve these problems. Now days, the network communication is becoming very important in the parallel programming environments of cluster computing, but most of the existing dynamic load balancing system are insufficiently supporting the programs that have network communication. This paper proposes a preemptive process migration system with the supporting of socket migration. Combined with the load balancing algorithm it can solve the above problem efficiently.An optimized total copy process migration algorithm is adopted by the process migration system proposed in this paper that can avoid transferring the dynamic linked library and code segment parts of the virtual address space of the migrating process. This optimization can greatly decrease the migration cost and the freeze time of the migrating process. Through the analysis of the implementation approaches and relevant systems, we find that the existing socket migration systems simply migrating the listening state socket from one node to another. The connection will failed if there is any client connecting to the original node with the original listening port. This will cause the whole program failure. We present the scene by describing the communication establishing steps of MPI programs using P4 library. We come to the conclusion that if don't solve this problem there will be no practical value of the existing socket migrating systems. The approach named SYN-CHANGE is proposed in this paper to solve the problem. The existing migration algorithm of the established socket is also been optimized by using the TCP timeout-retransmit mechanism that can improve the socket migration performance and decrease the overall freeze time. The preemptive process migration system proposed in this paper which named PPMM (Preemptive Process Migration Module) has been implemented as a Linux loadable kernel module. Because of need no modification of the operating system kernel, the system can improve the usability and maintainability of the process migration systems. The functional testes of the process migration system indicate that it can support socket migration perfectly. Through the performance measurement we find that the optimized process migration algorithm can greatly improve the performance.