A workload-adaptive and reconfigurable bus architecture for multicore processors

Deming Chen, Shoaib Akram, Alexandros Papakonstantinou, Rakesh Kumar

Research output: Contribution to journalArticlepeer-review


Interconnection networks for multicore processors are traditionally designed to serve a diversity of workloads. However, different workloads or even different execution phases of the same workload may benefit from different interconnect configurations. In this paper, we first motivate the need for workload-adaptive interconnection networks. Subsequently, we describe an interconnection network framework based on reconfigurable switches for use in medium-scale (up to 32 cores) shared memory multicore processors. Our cost-effective reconfigurable interconnection network is implemented on a traditional shared bus interconnect with snoopy-based coherence, and it enables improved multicore performance. The proposed interconnect architecture distributes the cores of the processor into clusters with reconfigurable logic between clusters to support workload-adaptive policies for inter-cluster communication. Our interconnection scheme is complemented by interconnect-aware scheduling and additional interconnect optimizations which help boost the performance of multiprogramming and multithreaded workloads. We provide experimental results that show that the overall throughput of multiprogramming workloads (consisting of two and four programs) can be improved by up to 60 with our configurable bus architecture. Similar gains can be achieved also for multithreaded applications as shown by further experiments. Finally, we present the performance sensitivity of the proposed interconnect architecture on shared memory bandwidth availability.

Original languageEnglish (US)
Article number205852
JournalInternational Journal of Reconfigurable Computing
StatePublished - 2010

ASJC Scopus subject areas

  • Hardware and Architecture


Dive into the research topics of 'A workload-adaptive and reconfigurable bus architecture for multicore processors'. Together they form a unique fingerprint.

Cite this