EyeQ: Practical Network Performance Isolation at the Edge

by Lakshminarayanan Subramanian

The basic question addressed in this paper is: "How do we provide the abstraction of network isolation across different competing tenants in a datacenter network, where each tenant gets the network abstraction of a dedicated switch?"  In other words, the datacenter network aims to provide bandwidth guarantees over short-time scales to thecompeting tenants in a datacenter. Achieving network-level isolation abstraction is a challenging problem, since the network resources are shared across multiple tenants and the datacenter network has limited knowledge about, and little control over, the volume of traffic from each tenant.  This paper presents EyeQ, a system that aims to provide fine-grained control for bandwidth partitioning across tenants in a datacenter network, and which aims to provide bandwidth guarantees to tenants. I believe that this paper solves a timely problem with growing traffic needs and contention levels at datacenters.

The key idea of their approach is to map the network isolation problem to a distributed and reactive congestion control protocol where they use rate meters in receivers that send congestion feedback to rate limiters at senders. Using a distributed feedback and sender based rate-limiting mechanism, the authors show that they can achieve predictable networkisolation over short time scales. EyeQ aims to guarantee a minimum bandwidth per each flow.  To achieve this, EyeQ uses a combination of three ideas: (a) the use of admission control at a flow level, using a hierarchical rate limiter at the sender side; (b) an aggregate tenant-level rate meter at the receiver, which provides control feedback to the sender; (c) the use of a variant of the Rate Control Protocol (RCP) at a per-flow level at the sender side, to provide per-flow max-min fairness within a pool of flows within a single tenant. One interesting result in this paper that EyeQ has a worst-case convergence time which is significantly lower than DCTCP and QCN, two related datacenter congestion control works. One should not view this paper as just a congestion control technique. The problem of providing predictable bandwidth guarantees turns out to be a much harder problem than achieving fairness and network sharing that conventional congestion control techniques aim to achieve.

There are several avenues to build upon this work. The bandwidth guarantees presented in this paper are probabilistic in nature. There is need for a more thorough theoretical analysis to quantify the nature of the guarantees, stability and convergence properties of the techniques presented in this paper. Achieving predictable network isolation in theface of misbehaving flows is inherently a hard problem, and this paper relies on the existence of a sender-level rate limiter to partially achieve this goal. The paper takes a specific design choice of using a rate control based approach to achieve predictable bandwidth guarantees. Given the large body of work on Internet Quality of Service (QoS), are there other techniques which could achieve a similar result or even improve the natureof achievable guarantees? To design a practical solution, one must also balance the tradeoff between deployability and achievable guarantees. Other avenues to consider in future work include exploring predictable guarantees for other network parameters like RTT and loss, and understanding the tradeoff between traffic volatility and achievable guarantees.