2015/02/10 by Daniel Gregorek, Gregorek, Daniel, Robert Schmidt +4
Computer Science · #Distributed #Distributed and Parallel Computing Systems #FOS: Computer and information sciences #Interconnection Networks and Systems #Parallel #Parallel Computing and Optimization Techniques #and Cluster Computing (cs.DC) #cs.DC
paper · pdf · doi:10.48550/arxiv.1502.02852
Presented at HIP3ES, 2015 (arXiv: 1501.03064)
arxiv created 2015/02/10 · openalex publication_date 2015/02/10 · arxiv updated 2015/02/11 · openalex created_date 2016/06/24 · openalex updated_date 2026/07/28
The increasing parallelism of many-core systems demands for efficient strategies for the run-time system management. Due to the large number of cores the management overhead has a rising impact to the overall system performance. This work analyzes a clustered infrastructure of dedicated hardware nodes to manage a homogeneous many-core system. The hardware nodes implement a message passing protocol and perform the task mapping and synchronization at run-time. To make meaningful mapping decisions, the global management nodes employ a workload status communication mechanism. This paper discusses the design-space of the dedicated infrastructure by means of task mapping use-cases and a parallel benchmark including application-interference. We evaluate the architecture in terms of application speedup and analyze the mechanism for the status communication. A comparison versus centralized and fully-distributed configurations demonstrates the reduction of the computation and communication management overhead for our approach.