7. Conclusion

A high-throughput computing policy attempts to efficiently utilize all of the resources in a cluster to generate faster overall execution. It is acceptable for jobs running on a workstation serving remote pages to execute more slowly, if other jobs benefit by achieving faster execution times through the use of network RAM, as long as throughput is maximized. Further work must be done to fully understand how the overhead inherent in providing network RAM affects running jobs on a workstation administering the service. With this knowledge, we will be able to adapt the load sharing policy to meet the efficiency requirements of the various load sharing goals. We have experimentally examined and compared job migrations and network RAM for sharing global cluster memory resources. Based on our experiments and analysis we have the following observations and conclusions: Providing a large memory space through remote paging , network RAM is particularly beneficial for large or data-intensive workloads where some jobs may not be migratable. However, the network RAM performance is heavily dependent on the cluster speed and the availability of the idle memory space in the cluster. Since load balancing is not considered, uneven job distributions may degrade the overall performance of cluster computing using network RAM. Dynamically migrating jobs by considering both the CPU and memory resources of the cluster, the load sharing policy using remote executions is particularly beneficial to data-intensive workloads where most jobs are migratable, and where each job fits in a memory space of a single workstation. The requirement of network speed by the remote-execution-based load sharing scheme is not as high as the network RAM. However , if the memory allocation of a job does not fit in any single workstation in the cluster, the additional memory requirement has to be satisfied by local disks, causing longer execution time. The improved load sharing scheme overcomes the limits and combines the advantages of the both schemes. We have shown that this scheme is effective for scal-able cluster computing. Memory allocations of jobs are generated by a Pareto distribution in the experiments presented in this paper. We have also run the simulations on the workloads with different memory demand distributions, and observed consistent performance results with that of the workloads by the Pareto memory demand distributions. The other distributions we have used for comparisons are uniform distribution, exponential distribution, and erlang distribution. Acknowledgements: We wish to thank Yanxia …