Accurate Contention-aware Scheduling Method on Clustered Many-core Platform

抄録

Embedded systems such as self-driving systems require a computing platform with high computing power and low power consumption. Multi-/many-core platforms definitely meet these requirements. However, for hard real-time applications, multiple demands on shared resources can hinder real-time performance. Memory is one of the resources that can most dramatically impair desired performance. Therefore, we addressed contentions induced by shared memory. The ability to predict contentions that may occur during memory access helps to reduce them. We improved the predictability of contentions by dividing tasks into the memory access phase and the execution phase using a Directed Acyclic Graph (DAG). Existing methods can make accurate contention estimations for one Compute Cluster (CC) of a clustered many-core processor. Our method is able to perform accurate contention estimations for multiple CCs, thereby doubling the scalability when contentions are taken into account. Using an Integer Linear Programming (ILP) formulation, we produced a static, non-preemptive, partitioned, and time-triggered schedule. We also conducted an experiment in order to minimize the makespan. The evaluation confirmed that our new method reduced the makespan by increasing the number of CCs.