Title: Eagle: Toward Scalable and Near-Optimal Network-Wide Sketch Deployment in Network Measurement
Speaker : Xiang Chen (Zhejiang University)
Scribe : Feiyan Ding (Xiamen University)
Authors:
Xiang Chen (Zhejiang University); Qingjiang Xiao (Southeast University); Hongyan Liu (Zhejiang University); Qun Huang (Peking University); Dong Zhang (Fuzhou University); Xuan Liu (Yangzhou University and Southeast University); Longbing Hu (ZTE Corporation); Haifeng Zhou, Chunming Wu, Kui Ren (Zhejiang University)
Introduction
Network measurement provides crucial information for understanding network status and managing operations. Traditional methods, like traffic sampling and traffic mirroring, have significant drawbacks: low accuracy for the former and high overhead for the latter. As a result, sketches have become essential for network measurement due to their low resource overhead and theoretical accuracy guarantees. Sketch deployment requires multi-objective optimization to meet diverse goals, adherence to various constraints, and efficient updates based on network changes. However, existing solutions have limitations: most rely on Mixed-Integer Linear Programming (MILP) solvers for optimal deployment decisions, which are time-consuming and difficult to scale for large networks. Other solutions use heuristic algorithms to improve scalability but at the cost of higher network resource usage and performance overhead.
Key idea and contribution:
The paper proposes a scalable and near-optimal framework for network-wide sketch deployment. The core idea involves using multi-objective optimization to address diverse deployment goals, managing heterogeneous constraints in programmable switches and networks, and ensuring efficient deployment through real-time updates of sketches in response to network changes.
Eagle first abstracts the network into a single large switch (OBS) and then decomposes the problem into two subproblems: Sketch → OBS and OBS → Network. The first subproblem uses stateful deployment to reuse deployment decisions from prior sketches for current sketches, and parallel deployment to parallelize the deployment of sketch components across different stages. For the second subproblem, it employs topology simplification to reduce the network to a model with only programmable switches, while preserving topology information without affecting deployment decisions. Finally, the results are integrated.
Evaluation
The paper evaluates the framework using various network topologies, including 5 WAN topologies and 5 data center topologies. Initially, it presents the overall timeliness of Eagle, demonstrating that Eagle reduces deployment time by two orders of magnitude. The paper then examines the timeliness of Eagle’s components, highlighting the performance impacts of Parallel Deployment and Topology Simplification. Finally, it discusses the optimality of Eagle, showing that the optimality gap is near zero. This result is significant because it means that Eagle maintains near-optimal performance.
Q1: In the Eagle framework, there’s an attempt to reuse one sketch when two are performing the same job to save resources. Could this approach lead to security issues? For instance, if two sketches need to be deployed from different tenants or applications, and the data is essentially on different private networks, might it contaminate the results?
A1: Eagle’s current version does not incorporate security considerations. Although this issue has not been addressed yet, there are plans to develop security techniques for integration into Eagle soon.
Q2: You mentioned the need for a solution that can respond to network dynamics in the motivational part of your talk. However, I didn’t see a detailed explanation of this in the technical sections. Could you elaborate on how Eagle addresses network dynamics?
A2: Eagle addresses network dynamics in two ways. First, new sketches are placed on ideal switches that were not used in previous deployments. Second, for optimal deployment of sketches, decisions are recalculated at runtime, and Eagle is used to expedite this process to ensure timely decisions can be made.
Personal thoughts
The paper introduces Eagle, a framework that tackles the scalability and near-optimality challenges in deploying network-wide sketches. I appreciate its innovative approach to decomposing the deployment problem into smaller, more manageable sub-problems, which is a key factor in achieving high efficiency. The use of the One Big Switch abstraction is a clever strategy that simplifies the deployment process. However, the paper could benefit from a deeper exploration of security implications, especially in multi-tenant environments where data isolation is crucial. An open question for future research could be how to integrate robust security measures into Eagle without sacrificing its scalability and efficiency.