[Paper Review] Towards Communication-Aware Robust Topologies

Chen Avin, Alexandr Hercules+2 more|arXiv (Cornell University)|May 19, 2017

Interconnection Networks and Systems15 references3 citations

TL;DR

This paper proposes a coding-based framework for designing sparse, robust network topologies that ensure expected path lengths are proportional to the entropy of communication patterns. By leveraging a continuous-discrete design and entropy-driven routing, the approach guarantees low average path lengths for frequently communicating nodes while maintaining resilience to edge failures and load balancing.

ABSTRACT

We currently witness the emergence of interesting new network topologies optimized towards the traffic matrices they serve, such as demand-aware datacenter interconnects (e.g., ProjecToR) and demand-aware overlay networks (e.g., SplayNets). This paper introduces a formal framework and approach to reason about and design such topologies. We leverage a connection between the communication frequency of two nodes and the path length between them in the network, which depends on the entropy of the communication matrix. Our main contribution is a novel robust, yet sparse, family of network topologies which guarantee an expected path length that is proportional to the entropy of the communication patterns.

Motivation & Objective

• To design network topologies that adapt to non-uniform traffic patterns, improving efficiency over traditional worst-case-optimized designs.
• To formalize a metric where path length is proportional to the entropy of the communication matrix, ensuring frequently communicating nodes are closer.
• To develop a robust, sparse topology family with provable guarantees on path length, load balancing, and failure resilience.
• To extend the continuous-discrete design framework to non-uniform probability distributions and enable greedy routing with forward/backward flexibility.
• To provide a theoretical foundation for demand-aware topologies that outperform existing approaches in entropy-aware performance and robustness.

Proposed method

• Utilizes a continuous-discrete design inspired by Naor and Wieder, enabling formal analysis in continuous space and discrete topology construction.
• Applies coding theory to map communication entropy to network topology structure, ensuring path length scales with entropy.
• Employs a greedy routing scheme that combines forward and backward routing to enhance flexibility and performance.
• Derives node degree and edge load expectations based on node activity (entropy), ensuring load balances with demand.
• Models edge failures probabilistically and evaluates resilience using TTL-based recovery mechanisms and weighted success rates.
• Uses simulation to compare Code-Based Routing (CBR) against non-code-based alternatives (CDR) across network size, load, and failure scenarios.

Experimental results

Research questions

• RQ1 Can network topologies be designed such that expected path length is proportional to the entropy of the communication matrix?
• RQ2 How can a continuous-discrete design be adapted to non-uniform traffic distributions for robust, sparse topologies?
• RQ3 To what extent does the proposed coding-based approach improve load balancing compared to conventional designs?
• RQ4 How does the topology perform under random edge failures, and can it maintain high delivery success with limited TTL?
• RQ5 Can greedy routing be enhanced by combining forward and backward routing to improve path efficiency and robustness?

Key findings

• The proposed Code-Based Routing (CBR) achieves expected path length (EPL) that grows proportionally with the entropy of the communication matrix, outperforming CDR which scales with log n.
• With 10% edge failure probability and TTL = 3× original path length, CBR delivers over 80% of messages, while CDR achieves less than 50% success rate.
• In CBR, node relay load is proportional to node activity (demand), ensuring balanced traffic distribution; in contrast, CDR shows no such correlation, leading to hotspot risks.
• For networks of size 100 to 1,000 nodes, CBR’s EPL increases with entropy, which can be significantly lower than the O(log n) bound of CDR.
• The framework enables robustness to failures through effective recovery mechanisms, with success rates improving as TTL increases beyond original path length.
• The coding-based approach enables both theoretical guarantees and practical performance improvements in path length, load balancing, and failure resilience.