Dynamic Load Sharing With Unknown Memory Demand of Jobs in Clustered Compute Farms

A compute farm is a pool of clustered workstations to provide high performance computing services for CPU-intensive, memory-intensive, and I/O active jobs. Existing load sharing schemes with memory considerations assume jobs’ memory demand sizes are known in advance or predictable based on users’ hints. This assumption can greatly simplify the designs and implementations of memory-centric schemes, but is not desirable in practice. Load sharing with unknown memory demand of jobs requires an insightful understanding of the effects of dynamic job interactions to memory systems and an effective usage of dynamic system information. In order to address this issue, we present three new results and contributions in this study. (1) Conducting Linux kernel instrumentations, we have collected different types of workload execution traces to quantitatively characterize job interactions, and modeled page fault behavior as a function of the overloaded memory sizes and the amount of jobs’ I/O activities. (2) Based on experimental results and collected dynamic system information, we have built a simulation model which accurately emulates the memory system operations and job migrations with virtual memory considerations. (3) We have proposed a memory-centric load sharing scheme and its variations to effectively process dynamic memory allocation demand of jobs, aiming at minimizing execution time of each individual job and maximizing distributed system throughput by dynamically migrating and remotely submitting jobs from heavily loaded machines to lightly loaded ones to eliminate or reduce page faults and to reduce the waiting time for CPU services. Conducting trace-driven simulations, we have examined these load sharing policies to show their effectiveness. Our study strongly suggests that load index be designed to comprehensively consider utilizing all the available resources including CPUs, global memory and network resources. This work is supported in part by the National Science Foundation under grants CCR-9400719, and CCR-9812187, and EIA9977030, by the Air Force Office of Scientific Research under grant AFOSR-95-1-0215, and by Sun Microsystems under grant EDUE-NAFO-980405.