End-point congestion control for high-throughput on-chip networks

Abstract

The advent of manycore processor architecure, on-chip network is an important aspect for high performance. Thus, there are huge requirements to achieve high network performance. One of the most important manycore processor architecture, General-Purpose computing on Graphics Processing Units (GPGPUs), it has distinguished features which are different front general architectures. In GPGPUs, the communication occurs between the cores and the memory controllers (MCs) not between the cores. Thus, there is a strong possibility to occur network congestion in the end-point (MCs). In this work, we propose an end-point congestion control mechanism for on-chip network. We base on a clumsy flow control (CFC) which is credit-based flow control between source to destination, not per-hop, and we present how to extend CFC to adjust the credits dynamically, refer as dynamic CFC. We propose three mechanisms to achieve it, i) gathering MCs queue occupancy, ii) dynamic credit self-tuning mechanism, and iii) fast credit return mechanism for optimizing. As a result, on average, dynamic CFC achieves 7% performance improvements compared to baseline buffered NoC, and 4% improvements compared to the best average performance of static CFC. Also, dynamic CFC performance achieves the best performance with the optimal number of credits in static CFC at each workload.