TY - GEN
T1 - A Virtual-Channel Free Mapping for Application-Specific On-Chip Torus Networks
AU - Matsutani, Hiroki
AU - Koibuchi, Michihiro
AU - Amano, Hideharu
N1 - Funding Information:
This work was supported by Joint Research Fund, “Network-on-Chip Architecture,” National Institute of Informatics.
Publisher Copyright:
© 19th International Conference on Parallel and Distributed Computing Systems 2006, PDCS 2006. All rights reserved.
PY - 2006
Y1 - 2006
N2 - Various types of Networks-on-Chips (NoCs) have been employed light-weight routers compared with those in parallel computers, and a virtual-channel mechanism, which requires additional logic and pipeline stages, is one of the crucial factors for a low cost implementation of an NoC router in the case of simple tile-based architectures. Although a torus network, which exploits wrap-around channels, achieves higher throughput and lower latency than a same-sized mesh, a virtual-channel mechanism is usually required to avoid deadlocks in tori with dimension-order routing. In this paper, we propose a scheme to remove virtual channels in tori by accomplishing the following steps: 1) providing a mechanism which allows wraparound channels to be individually disabled in each router, 2) a task mapping strategy that carefully assigns tasks to a tori, so that as many wrap-around channels as possible are exploited without introducing deadlocks or performance degradation. Additionally, we extend this strategy to avoid deadlocks when application traffic patterns are unknown or incompletely analyzed. Although the proposed mapping does not use virtual channels, it achieves almost the same performance as conventional mapping in tori in ten traces. Moreover, the hardware amount of the proposed router can be decreased to 52.4% of a conventional router providing two virtual channels for tori.
AB - Various types of Networks-on-Chips (NoCs) have been employed light-weight routers compared with those in parallel computers, and a virtual-channel mechanism, which requires additional logic and pipeline stages, is one of the crucial factors for a low cost implementation of an NoC router in the case of simple tile-based architectures. Although a torus network, which exploits wrap-around channels, achieves higher throughput and lower latency than a same-sized mesh, a virtual-channel mechanism is usually required to avoid deadlocks in tori with dimension-order routing. In this paper, we propose a scheme to remove virtual channels in tori by accomplishing the following steps: 1) providing a mechanism which allows wraparound channels to be individually disabled in each router, 2) a task mapping strategy that carefully assigns tasks to a tori, so that as many wrap-around channels as possible are exploited without introducing deadlocks or performance degradation. Additionally, we extend this strategy to avoid deadlocks when application traffic patterns are unknown or incompletely analyzed. Although the proposed mapping does not use virtual channels, it achieves almost the same performance as conventional mapping in tori in ten traces. Moreover, the hardware amount of the proposed router can be decreased to 52.4% of a conventional router providing two virtual channels for tori.
UR - http://www.scopus.com/inward/record.url?scp=84947789914&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84947789914&partnerID=8YFLogxK
M3 - Conference contribution
AN - SCOPUS:84947789914
T3 - 19th International Conference on Parallel and Distributed Computing Systems 2006, PDCS 2006
SP - 24
EP - 31
BT - 19th International Conference on Parallel and Distributed Computing Systems 2006, PDCS 2006
PB - International Society for Computers and Their Applications (ISCA)
T2 - 19th International Conference on Parallel and Distributed Computing Systems, PDCS 2006
Y2 - 20 September 2006 through 22 September 2006
ER -