A Control-Theoretic Approach to Auto-Tuning Dynamic Analysis for Distributed Services
Chandan Dhal, Xiaoqin Fu, Haipeng Cai · 2023
Traditional dynamic dependence analysis approaches have limited utilities for continuously running distributed systems (i.e., distributed services) because of their low cost-effectiveness. A recent technique, SEADS, was developed to improve the cost-effectiveness by adjusting analysis configurations on the fly using a general Q-learning algorithm. However, SEADS is unable to utilize the user budget as far as needed for pushing up precision. To overcome this problem, we propose CADAS, an adaptive dynamic dependency analysis framework for distributed services. To realize the adaptation, we are exploring a control-theoretical method which uses a feedback mechanism to predict optimal analysis configurations. Then, we evaluated CADAS against six real-world Java distributed services. We compared CADAS against SEADS as the baseline and show that CADAS outperforms the baseline in both precision and budget utilization. Our results suggest a new door opening for future research on adaptive dynamic program analysis.