Performance Evaluation of Self-Similar GPS Scheduling: A Knowledge-Driven Dual-Task Deep Learning Approach

Generalized Processor Sharing (GPS) under self-similar traffic is widely used for guiding resource allocation in modern communication networks where performance evaluation, i.e., average waiting queue length and average waiting queue delay, is the key step. Due to its high complexity and unknown inherent relations, theoretically attaining the accurate performance of each queue remains unsolved. Simulation can obtain accurate performance but the time cost is unaffordable. In this paper, we address the performance evaluation for multi-queue GPS under self-similar traffic problem and propose a knowledge-driven dual-task learning approach, namely DualTask_GPS. Firstly, DualTask_GPS utilizes attention mechanisms to model the inherent relations in GPS based on augmented queue-theory-based knowledge-driven features and a gate-based server-queue module to adaptively fuse the learned information of the server and queues to obtain the intermediate capacity corresponding to each queue for the final performance prediction. To capture more reasonable representations, we jointly learn to predict the average waiting queue length and average waiting queue delay of a GPS system with posterior constraints by prior knowledge. To evaluate the performance of the DualTask_GPS model, we conduct experiments on the three widely acknowledged GPS settings. The results show that our proposed approach is effective regarding various metrics on both tasks.

Download paper here