Criticality Avoidance A new paradigm for congestion control based on science of phase transition Siun-Chuon Mau, Akshay Vashist, Alexander Poylisher, Ritu Chadha, and Cho-Yu Jason Chiang Telcordia Technologies, Inc. Piscataway, NJ, USA {smau, vahist, sher, chadha, chiang}@research.telcordia.com Abstract—Network QoS control is generally difficult due to the complexity, dynamism, and limited measurability of networks. As an alternative, we seek a network phenomenon that is simple, universal and consequential to control. The result is a framework for proactive dynamic network congestion control that is based on the science of continuous phase transition. Key beneficial properties of continuous phase transition are its early onset warning signs and universality. The former allows the detection of proximity to congestion before its occurrence; while the latter implies that any criticality-based network control would likely be insensitive to network details and, in particular, not require any a-priori knowledge of the values of critical loads. Preliminary experimental results demonstrating these promises are presented. Keywords-congestion; QoS; quality of service; phase transition; critciality; onset detection; control. I. INTRODUCTION Quality-of-Service (QoS) control for communications networks is generally difficult due to the complexity, dynamism, and limited measurability of the network. These attributes arise because of the existence of a large number of interacting functional and architectural components that operate on different time scale, are opaque to each other, and yet can affect the same QoS metrics. For example, rate allocation can involve end-to-end congestion control, local scheduling, per-hop resource allocation, routing, and even human network operators [1], [8]. As a consequence, QoS controls tend to be ad hoc and incomplete. The most well-known effort to reorganize and simplify network architecture evolves from the application of optimization decomposition to reverse engineer TCP [10], [11], [12], into using it as a layering principle [1], with numerous applications, such as traffic management at carriers [8] and a new version of TCP for high-speed long-latency networks [13]. In our view, a key benefit of this approach is that it allows systematic derivations of QoS controls that are grounded in principle and are less likely to suffer from the usual shortcomings of ad hoc, case-by case QoS controls. Against this backdrop, the goal of this article is to suggest, for the purpose of controlling network congestion, another opportunity to simplify network control and reap similar (and additional, see below) benefits by viewing network congestion from the perspective of the well established science of continuous phase transition (CPT) or continuous criticality. We will offer evidence that network congestion is indeed a CPT and that the properties of CPT afford criticality-based congestion control the aforementioned benefits. The study of network congestion as a phase-transition phenomenon exists in the physics community. In [15] and [16], simulation results on wired networks show that the congestive transition is of the continuous variety. However, their scope was to understand congestive criticality, rather than to leveraging its properties for network control, as we desire to. The knowledge that congestive phase transition is continuous is emphasized here, because the behaviors of warning signs for continuous criticalities are well understood, while those for discontinuous criticalities are not. Incidentally, they are also called second-order and first-order phase transitions, respectively. Certain shared properties of CPTs are particularly valuable due to their implications for network control. First, a system with CPT exhibits measurable, advanced warning signs, as it approaches the critical, or transition, point [18], [19], making it possible to proactively (vs. reactively) avoid the CPT. Such warning signs in complex dynamical systems including ecological, climatic, and financial are discussed in [17]. Second, CPTs are universal [18], [19] in the sense that all systems progressively “behave the same” (upon scaling) as a transition point is approached. (See Figure 1. and 2.) In particular, this implies that no a priori knowledge of the value of transition point is required for the control of a CPT. An issue congestion-control designers might be concerned with is how much network details matter. The answer is very minimal, at least for the explicit control of congestive criticality; this is a major simplification. While network congestion is well studied, its warning signs and the value of its critical load via network information theory remain poorly understood. Employing the science of CPT allows for a systematic understanding of the former and renders the latter less important for network-congestion control. The main contributions of this article are to draw attention to the significant potential of the CPT-view of network congestion and to report initial simulation results demonstrating the predictability and measurability of network congestion warning signs. This work is solely supported internally by Telcordia Technologies, Inc. A US patent application is filed based on this work.