Cache-Aware Allocation of Parallel Jobs on Multi-cores based on Learned Recency

Abstract

Scheduling of tasks on multi-and many-cores benefits significantly from the efficient use of caches. Most previous approaches use the static analysis of software in the context of the processing hardware to derive fixed allocations of software to the cache. However, there are many issues with this approach in terms of pessimism, scalability, analysis complexity, maintenance cost, etc. Furthermore, with ever more complex functionalities being implemented in the system, it becomes nearly impracticable to use static analysis for deriving cache-aware scheduling methods. This paper focuses on a dynamic approach to maximise the throughput of multi-core systems by benefiting from the cache based on empirical assessments. The principal contribution is a novel cache-aware allocation for parallel jobs that are organised as directed acyclic graphs (DAGs). Instead of allocating instruction and data blocks to caches, the proposed allocation operates at a higher abstraction level that allocates jobs to cores, based on the guidance of a predictive model that approximates the execution time of jobs with caching effects taken into account. An implementation of the predictive model is constructed to demonstrate that the execution time approximations can be effectively obtained. The experimental results, including a real-world case study, prove the concept of the proposed cacheaware allocation approach and demonstrate its effectiveness over the state-of-the-art. CCS CONCEPTS• Computer systems organization → Real-time system architecture; Processors and memory architectures.