Until now little research has addressed the problem of on line scheduling of sporadic parallel tasks with hard deadlines in partitionable multiprocessor systems In this paper we present two on-line scheduling algorithms - Buddy/RTand Stacking, for such environments Both the algorithms either guarantee or reject a task at the time of its arrival Buddy/RT is a straight-forward extension of the well-known Buddy strategy to the real-time environment, while Stacking is a more sophisticated algorithm based on the lessons learned from Buddy/RT The underlying concept behind the Stacking algorithm is to reduce fragmentation by 'stacking' equal-sized jobs in the time dimension The Stacking algorithm is found to perform significantly better than Buddy/RT over a wide range of workloads, even though both the algorithms have the same time complexity

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 allocation). 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 Scan, 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.>

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 receiver-initiated, sender-initiated, or symmetrically-initiated. The relative performance of these classes of algorithms has been shown to depend on the system workload. We present two adaptive location policies for global scheduling in distributed systems. These location policies are general, and can be used in conjunction with many existing transfer policies. By adapting to the system workload, the proposed policies capture the advantages of both sender-initiated and receiver-initiated policies. In addition, by adaptively directing their search activities toward the nodes that are most likely to be suitable counterparts in task transfers, the proposed policies provide short transfer latency and low overhead, and more important, high probability of finding a suitable counterpart if one exists. These properties allow these policies to deliver good performance over a very wide range of system operating conditions. The proposed policies are compared with nonadaptive policies, and are shown to considerably improve performance and to avoid causing system instability.>

A key issue in distributed scheduling is selecting appropriate jobs to transfer. A job selection policy that considers the diversity of job behaviors is proposed. A mechanism used in artificial neural networks, called weight climbing, is employed. Using this mechanism, a distributed scheduler can learn the behavior of a job from its past executions and make a correct prediction about whether transferring the job is worthwhile. A scheduler using the proposed job selection policy has been implemented and experimental results show that it is able to learn job behaviors fast, make decisions accurately and adjust itself promptly when system configuration or program behaviors are changed. In addition, the selection policy introduces only negligible time and space overhead.>

The justification, design, and performance of the Stealth distributed scheduler is discussed. The goal of Stealth is to exploit the unused computing capacity of a workstation-based distributed system (WDS) without undermining the predictability in quality of service that a WDS provides to workstation owners. It is shown that the liberal approach taken by the Stealth distributed scheduler is a promising method of exploiting the vast quantity of unused computing capacity typically present in a WDS, while preserving predictability of service for workstation owners.>

The roles of processor allocation and job scheduling in achieving good performance on hypercube computers are compared. It is shown that the choice of job scheduling discipline has a dramatic effect on performance. A family of scheduling disciplines, called Scan, with particular performance advantages is proposed. Furthermore, it is shown that if Scan scheduling is used, the choice of processor allocation strategy has negligible effect on performance. As a result, complex allocation strategies can be replaced by a simple O(n) strategy.>

Two location policies that, by adapting to the system load, capture the advantages of receiver-initiated, sender-initiated, and symmetrically initiated algorithms are presented. 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. By means of simulation, two representative algorithms making use of these adaptive location policies are shown to be stable and to improve performance significantly relative to nonadaptive policies.>

The authors consider whether the addition of a migration facility to a distributed scheduler already capable of placement can significantly improve performance. They examine performance over a broad range of workload characteristics and file system structures. They find that while placement alone is capable of large improvement in performance, the addition of migration can achieve considerable additional improvement.>