DRACON : a dedicated hardware infrastructure for scalable run-time management on many-core systems

Abstract

Many-core architectures integrate a large number of comparatively small processing cores into a single chip. However, the high degree of parallelism increases the run-time resource management complexity and overhead. The employment of dedicated hardware enhancements potentially enables a high quality of the resource management while management overhead is mitigated. To exploit the potential of hardware enhancements, we propose a dedicated infrastructure for run-time resource management on homogeneous MIMD many-core processors. For hardware enhanced resource management, a scalable and cluster-based system architecture is implemented. The resulting architecture (DRACON) utilizes message passing based communication, the dedicated infrastructure and hardware accelerators for resource management. A comprehensive evaluation for DRACON and reference architectures is performed using a transaction level simulation framework and dynamic task management as a use case. As benchmarks, synthetic models and task graph models of real-world applications are applied. The results reveal the limited scalability of classical architectures for resource management on many-cores. It is therefore necessary to apply cluster-based or moderately distributed architectures for many-core resource management. Further, the results demonstrate a significant performance improvement for the DRACON architecture at a number of hundreds of processing cores. Our evaluations show that DRACON generally outperforms software-only run-time management on many-core and achieves a performance improvement of up to 15.21% for single-program and more than 6% for mixed workloads.