A Unified Job Scheduler for Optimization of Different System Performance Metrics

Abstract

Internet-of-Things-enabled frameworks have eased the development of complex systems, but they throw a significant challenge for efficient resource utilization, thereby improving the system performance. An intelligent scheduler is essential for managing the resources and allocating the same resources to different requests or tasks. This work proposes a generic methodology to optimize system performance metrics such as throughput, utilization, and reward achieved. We present an integer linear programming formulation of the problem to find an optimal solution. We present offline heuristic methods to quickly find reasonable solutions, given the intractable nature of the problem. These heuristics yield promising outcomes, with deviations from optimal solutions below 20% in scenarios with task overlap and high utilization. In scenarios with minimal overlap and utilization, deviations remain under 10%. However, as variables and constraints increase in ILP, the demand for time and memory resources rises substantially. We conduct a comparative analysis of heuristic performance across various scenarios and large test cases. Additionally, we extend our methods to handle resources in online mode, presenting an extensive comparative study with encouraging results.