[Paper Review] Online Dynamic B-Matching With Applications to Reconfigurable Datacenter Networks

Marcin Bieńkowski, David Fuchssteiner+2 more|arXiv (Cornell University)|Jun 18, 2020

Optimization and Search Problems69 references4 citations

TL;DR

This paper introduces an online dynamic b-matching algorithm for reconfigurable datacenter networks, where nodes can maintain up to b reconfigurable optical links to optimize routing based on real-time traffic. The proposed algorithm achieves O(b)-competitiveness, which is asymptotically optimal, and is validated through trace-driven simulations on real and synthetic workloads.

ABSTRACT

This paper initiates the study of online algorithms for the maximum weight $b$-matching problem, a generalization of maximum weight matching where each node has at most $b \geq 1$ adjacent matching edges. The problem is motivated by emerging optical technologies which allow to enhance datacenter networks with reconfigurable matchings, providing direct connectivity between frequently communicating racks. These additional links may improve network performance, by leveraging spatial and temporal structure in the workload. We show that the underlying algorithmic problem features an intriguing connection to online paging (a.k.a. caching), but introduces a novel challenge. Our main contribution is an online algorithm which is $O(b)$-competitive; we also prove that this is asymptotically optimal. We complement our theoretical results with extensive trace-driven simulations, based on real-world datacenter workloads as well as synthetic traffic traces.

Motivation & Objective

• To address the challenge of dynamically optimizing reconfigurable optical links in datacenter networks to reduce bandwidth tax and reconfiguration costs.
• To model the problem as an online dynamic b-matching problem, where each node can connect to at most b other nodes via reconfigurable links.
• To design an online algorithm that balances routing efficiency and reconfiguration cost in real time, under partial knowledge of future requests.
• To establish theoretical competitive ratios and prove tightness of the performance bound.
• To evaluate the algorithm’s practical performance using real-world and synthetic datacenter traffic traces.

Proposed method

• Formalizes the problem as an online dynamic b-matching problem on a static weighted network, where matching edges represent reconfigurable optical links.
• Models routing cost as the shortest path length in the fixed network if no matching edge exists, and zero cost if a direct matching edge is available.
• Introduces a novel online algorithm that maintains a b-matching over time, adapting to incoming communication requests without prior knowledge of future traffic.
• Establishes a connection to online paging (caching) to guide algorithm design, though with new structural challenges due to the matching constraints.
• Uses competitive analysis to evaluate performance, comparing the online algorithm’s cost to that of an optimal offline adversary.
• Employs trace-driven simulations using real-world datacenter workloads and synthetic traffic to evaluate practical performance.

Experimental results

Research questions

• RQ1 What is the optimal competitive ratio achievable for online dynamic b-matching in reconfigurable datacenter networks?
• RQ2 How can online algorithms be designed to balance routing cost and reconfiguration cost in a dynamic setting with partial request information?
• RQ3 What is the theoretical limit of performance for online b-matching, and can this bound be achieved?
• RQ4 How does the proposed algorithm perform under realistic datacenter traffic patterns compared to baseline strategies?
• RQ5 Can the connection to online paging be leveraged to design efficient and competitive algorithms for b-matching?

Key findings

• The proposed online algorithm achieves an O(b)-competitive ratio, which is asymptotically optimal as proven by a matching lower bound.
• The algorithm is simple to implement and exhibits low runtime, making it suitable for practical deployment in datacenter environments.
• Trace-driven simulations show strong performance across diverse real-world and synthetic workloads, confirming practical viability.
• The problem exhibits a novel connection to online paging, but introduces additional structural constraints that complicate direct adaptation of existing caching techniques.
• The study identifies that randomized algorithms may be more challenging due to dependencies introduced by the matching constraints and the reconfiguration cost model.
• The work opens new research directions in dynamic network optimization, particularly for distributed implementations and integration with specific optical technologies and topologies like Clos networks.