Search Results

Task-to-processor assignment is an essential aspect of configuring real-time, distributed systems, since an improper assignment can adversely affect latency. Model-based, heuristic, and data-driven approaches have been proposed to solve the task-to-processor assignment problem. However, model-based and heuristic approaches require revision if the system changes, and data-driven approaches require training on a lot of data and setting nonintuitive hyper-parameters. We explore a hybrid approach which takes both a system description and data: we use inductive logic programming in an active learning algorithm to search for assignments which satisfy a real-time requirement. By using both domain knowledge and data, the system finds solutions quickly, and changes are not required when using the tool on different systems. Furthermore, the output is a human-readable description of a set of predicted satisfactory assignments. Readable solution sets are useful for analyzing the system, since we can easily compare solution sets across different setups. We evaluate our approach on real systems with mixed-critical network flows. We show that task-to-processor assignment can significantly influence latency by comparing optimal fixed assignments to the default Linux scheduler. We show that our approach finds assignments that are within 10% of optimal with up to 10× fewer system tests, compared to random search. Our algorithm also performs favorably to load balancing and neural network baselines.