B-RPM: An Efficient One-to-Many Communication Framework for On-Chip Networks

Date

2012-10-19

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

The prevalence of multicore architectures has accentuated the need for scalable on-chip communication media. Various parallel applications and programming paradigms use a mix of unicast (one-to-one) and multicast (one-to-many) to maintain data coherence and consistency. Providing efficient support for these communication patterns becomes a critical design point for on-chip networks (OCN). High performance on-chip networks design advocates balanced traffic across the whole network, which makes adaptive routing appealing. Adaptive routing explores the path diversity of the network, increases throughput, and reduces network latency compared with oblivious routing.

In this work, we propose an adaptive multicast routing, Balanced Recursive Partitioning Multicast (B-RPM), to achieve balanced one-to-many on-chip communication. The algorithm derives its functionality from previously proposed algorithm Recursive Partitioning Multicast (RPM). Unlike RPM which uses fixed set of directional priorities and position of destination nodes, B-RPM replicates packet based on the local congestion information and position of destination nodes with respect to current node. B-RPM employs a new deadlock avoidance technique Dynamically Sized Virtual Networks (DSVN). Built upon the traditional virtual networks, DSVN dynamically allocates the network resources to different VNs according to the run-time traffic status, which delivers better resources utilization. We also propose a new scheme for representing multiple destinations in packet head. The scheme works simply by differentiating multicast and unicast packets. The algorithm combined with dynamically sized virtual networks enables us to improve network performance at high load on average by 20% (up to 50%) and saturation throughput of network on average by 10% (up to 18%) over the most recent multicast algorithm. Also the new header representation scheme enables us to save 24% of dynamic link power.

Description

Citation