Modeling NoC traffic locality and energy consumption with rent's communication probability distribution

Abstract

In systems on chip, the energy consumed by the Network on Chip (NoC) depends heavily on the network traffic pattern. The higher the communication locality, the lower the energy consumption will be. In this paper, we use the Communication Probability Distribution (CPD) to model communication locality and energy consumption in NoC. Firstly, based on recent results showing that communication patterns of many parallel applications follow Rent's rule [6], we propose a Rent's rule traffic generator. In this method, the probability of communication between cores is derived directly from Rent's rule, which results in CPDs displaying high locality. Next, we provide a model for predicting NoC energy consumption based on the CPD. The model was tested on two NoC systems and several workloads, including Rent's rule traffic, and obtained accurate results when compared to simulations. The results also show that Rent's rule traffic has lower energy consumption than commonly used synthetic workloads, due to its higher communication locality. Finally, we exploit the tunability of our traffic generator to study applications with different locality, analyzing the impact of the Rent's exponent on energy consumption.

The Hamiltonian-based odd–even turn model for maximally adaptive routing in 2D mesh networks-on-chip

“…Therefore, the locality of tasks is required in NoCs, analogous to the classical logic placement in circuits [22]. Extending the concept of connection locality in circuits to communication locality among the cores was the motivation for Greenfield et al [22] to apply the principle of Rent's rule to the analysis of NoCs [23].…”

“…Hence, we have implemented a Rentian traffic generator in order to assess the efficiency of the proposed method under real traffic distributions. The implementation is based on the Communication Probability Distribution (CPD) introduced in [23]. CPD is the probability that a processor sends packets to another processor at a particular distance in the network: the farther the distance, the smaller the probability.…”

“…CPD is the probability that a processor sends packets to another processor at a particular distance in the network: the farther the distance, the smaller the probability. The formula for CPD and implementation details are described in detail in [23].…”

Design of TSV-Sharing Topologies for Cost-Effective 3D Networks-on-Chip

“…The performance of the proposed models is evaluated under the uniform, hotspot 20%, and Rentian [17] traffic patterns. For the Rentian traffic, the Communication Probability Distribution (CPD) determines the locality of the traffic in the network [17] which is set to 70% in our simulations.…”

“…For the Rentian traffic, the Communication Probability Distribution (CPD) determines the locality of the traffic in the network [17] which is set to 70% in our simulations. The average latency curves of five topologies as a function of the network's request rate are plotted in Fig.…”