A Distributed Look-Ahead Workload Assignment Algorithm for Interdependent Tasks

Autonomous Decentralized Systems concurrently work on different types of jobs. From the sys-tem's point of view, every job consists of interdependent tasks (sreps). Steps are characterizedby their service requirements. It is the problem of load sharing to increase the system throughput by reducing contention between steps that seek access to the same resources. However, traditional approaches to dynamic decentralized , controlled load balancing disregard job-internal step dependencies and the need to support the execution of jobs with different future service requirements in distributed systems. In this paper, a dynamic decentralized look-ahead workload assignment algorithm is proposed together with a cooperation protocol for exchanging the required information. The goal is to utilize partial knowledge about the internal job structure concerning future service requirements and system state information for dynamically arranging schedules so that some jobs can take advantage of the inevitable waiting times of others. This can significantly increase system performance by reducing job waiting times. The evaluation of the algorithm under various load conditions is based on simulation studies. Waiting-time reductions in sample configurations are up to 90 percent for single job types at the expense of only slightly worse response times for other job types, compared with systems not applying the algorithm that prove the algorithm's success.