Design and Analysis of an Incast Mitigation System

Abstract

In this work, we study the long-standing incast congestion problem, which is predominant in modern communication networks such as data center networks. The incast congestion is notorious for being the main source of increased delays in modern network applications. For instance, incast is the main performance bottleneck for partition-aggregate processing workflows in data center networks. Typically, these workflows involve multiple data sources and a shared queue with limited buffering capacity. We propose a system based on active queue management, which aims to mitigate the effects of incast and provide a detailed stability analysis of the proposed system. The proposed system leverages a hysteresis controller that switches between different data rates based on the queue length. In this work, we present the system model, including the control loop and the switching control law used by the switch. We present a novel theoretical system analysis that leverages Lyapunov functions and Lagrange stability, showing the system's stability. We also support our findings via simulation analysis in NS2.