Investigating the Viability of Maximum Flexibility Selection Function in Bufferless 2D Meshes

Abstract

Bufferless NoCs have emerged as a solution to reduce power and area by eliminating buffers used for routing. Such networks handle contention using packet dropping or deflection. In this paper, we study the effect of MaxFlex selection function on 2D bufferless meshes for both a fixed and a variable step size. For fixed step size, we perform an analytical study for the effect of using MaxFlex with different step size on the performance of 2D bufferless meshes. The analysis indicates that, as the step size increases the traffic in the central part of the network bisection relaxes. Simulation results show that, both average packet latency and average deflection count decrease as the step size used increases. Additionally, over different sizes of meshes, the results show that the network performs best if the step size is equal 60--80% of the mesh dimension. Then, we consider using variable step size in which a packet is routed using a step size dependent on the Manhattan distance, d, between the source and destination. Simulation results show that, using MaxFlex, a step size of 60% of the distance d enhances the packet latency over using fixed step size, straight line selection function and random productive port selection function by around 29%, 97% and 99% respectively.