EE Seminar: Efficient Communication in Manycore Processors using Deflection Routing
~~Speaker: Gadi Oxman
Ph.D. student under the supervision of Prof. Shlomo Weiss
Wednesday, July 6th, 2016 at 15:00
Room 011, Kitot Bldg., Faculty of Engineering
Efficient Communication in Manycore Processors using Deflection Routing
Abstract
Scaling technology limitations shifted the focus in the computer architecture field from complex single threaded cores with ever increasing operating frequencies, to integration of many parallel simpler cores on the same silicon die. A key component that facilitates communication among the cores is the network-on-chip (NoC). The NoC embeds many interconnected routers with the cores in the same chip. Since the router is replicated many times, it is important to improve its performance and power consumption.
A low cost NoC routing scheme that we focus on in this work is deflection routing (also called hot potato routing), originally proposed for off-chip networks. Deflection routing can be bufferless, thus reducing power consumption by eliminating the router's buffers altogether. However, this comes at a network performance cost, especially under heavy load. We propose the use of buffers in deflection routing to improve performance and power efficiency, and show that substantial improvement can be achieved even with a small number of buffers.
Deflection routing is inherently adaptive, and there are multiple routes by which a packet can reach its destination. We propose prioritization schemes that improve performance and reduce network congestion. We also improve deflection routing performance on diagonal, irregular, and hierarchical mesh topologies.
We describe a new cycle based NoC simulation tool that was developed during our research. The simulator can simulate the performance, power and energy of NoCs using deflection routing. It supports both stand-alone simulation using synthetic traffic, as well as co-simulation along with a system simulator, facilitating closed-loop simulation of NoC traffic from application benchmarks.
