Managing computing resources in a hypercube entails two steps. First, a job must be chosen to execute from among those waiting (job scheduling). Next a particular subcube within the hypercube must be allocated to that job (processor allocaabn). Whereas processor allocation has been well studied, job scheduling has been largely neglected. The goal of this paper is to compare the roles of processor allocation and job scheduling in achieving good performance on hypercube computers.We show that job scheduling has far more impact on performance than does processor allocation. We propose a new family of scheduling disciplines, called Scun, that have particular performance advantages. We show that performance problems that cannot be resolved through careful processor allocation can be solved by using Scan job-scheduling disciplines. Although the Scan disciplines carry far less overhead than is incurred by even the simplest processor allocation strategies, they are far more able to improve performance than even the most sophisticated strategies. Furthermore, when Scan disciplines are used, the abilities of sophisticated processor allocation strategies to further improve performance are limited to negligible levels. Consequently, a simple O(n) allocation strategy can be used in place of these complex strategies.
Numerous load distributing algorithms have been proposed over the past several years, with widely varying characteristics. While some of these algorithms rely solely on non-preemptive process placement, others make use of preemptive process migration. Because the state of a process becomes considerably more complex after it begins execution, the mechanism necessary for migration is correspondingly more complex than that for placement, and may incur significantly greater resource overhead. In light of this complexity, as well as the consequent implementation expense, we consider whether the addition of a migration facility to a distributed scheduler already capable of placement can significantly improve performance. We examine performance over a broad range of workload characteristics and file system stNctures. We Jind that, while placement alone is capable of large improvement in performance, the addition of migration can achieve considerable additional improvement.
Two important components of a global scheduling algorithm are its transfer policy and its location policy. While the transfer policy determines whether a task should be transferred, the location policy determines where it should be transferred. Based on their location policies, global scheduling algorithms can be broadly classified as receiverinitiated, sender-initiated, or symmetrically-initiated. Unfortunately, choosing from among these classes of algorithms is difficult, because how well each performs relative to the others and the effect each has on system stability has been shown to depend on the system load. In this paper, we present two new location policies that, by adapting to the system load, capture the advantages of all three classes of algorithms. A key feature of these location policies is that they are general, and can be used in conjunction with a broad range of existing transfer policies. Using simulation, two representative algorithms making use of these adaptive location policies are shown to be stable and to significantly improve performance relative to non-adaptive policies.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.