Stability-Guaranteed Optimization for Real-Time Task Offloading in Edge Computing Systems

Abstract

The importance of edge computing for applications with hard deadlines is on the rise. For these applications, offloading a task isn't just a matter of where the task is to be executed - the time and resources needed for the offloading decision itself can also eat into the slack time before the deadline. To solve this challenge, this paper presents Stable-Guaranteed Adaptive Optimization (SGAO), a real-time optimization engine that takes the resources needed for the optimization as a constrained system resource. These features are key to safe, always-on operations.SGAO sets the maximum number of iterations per task based on the slack available in the deadline and the estimated cost of one iteration of the optimization. This is complemented by a stability-based stop criterion, so that at some point, the change in consecutive objective values does not justify further computation. Hence, the proposed approach ties the urgency of a real-time task to the maximum computational effort that may be spent on it, with a simple deadline-aware iteration limit and an explicit limit on the variation in the objective. A controlled synthetic benchmark compares eight different approaches at five workload levels. For each workload level, 2,000 tasks are simulated with ten independent seeds. The experiment results demonstrate the potential of SGAO to greatly reduce deadline violations relative to fixed number-of-iterations SGAO and other optimisation-based baselines, while incurring drastically reduced decision time relative to an optimizer with a relatively high fixed iteration limit. The experiment results do not, however, show SGAO outperforming every other approach on all metrics; in synthetic experiments, Greedy-Latency demonstrates a slightly lower deadline-miss rate while incurring lower latency. Far from ignoring this behaviour, the results are presented explicitly as they define the practical scope of the approach. In doing so, SGAO is not defined as a learning-based or heuristic-based alternative to Lyapunov-based, learning-based, or heuristic task offloading approaches, but as a real-time optimisation-control layer which minimizes offload optimization effort subject to deadline constraints.