[Paper Review] A Flat and Scalable Data Center Network Topology Based on De Bruijn Graphs

Frank Dürr|arXiv (Cornell University)|Oct 11, 2016

Interconnection Networks and Systems5 citations

TL;DR

This paper proposes a flat, scalable data center network topology using De Bruijn graphs to eliminate hierarchical switches and reduce hardware cost. By leveraging prefix matching in TCAM for efficient, low-overhead routing with minimal forwarding table entries, the approach achieves logarithmic path lengths and enables deployment on commodity switches—offering a viable alternative to Leaf-Spine topologies despite a 25% throughput reduction compared to ECMP in Clos networks.

ABSTRACT

Due to the requirement of hosting tens of thousands of hosts in today's data centers, data center networks strive for scalability and high throughput on the one hand. On the other hand, the cost for networking hardware should be minimized. Consequently, the number and complexity (e.g. TCAM size) of switches has to be minimized. These requirements led to network topologies like Clos and Leaf-Spine networks only requiring a shallow hierarchy of switches---two levels for Leaf-Spine networks. The drawback of these topologies is that switches at higher levels like Spine switches need a high port density and, thus, are expensive and limit the scalability of the network. In this paper, we propose a data center network topology based on De Bruijn graphs completely avoiding a switch hierarchy and implementing a flat network topology of top-of-rack switches instead. This topology guarantees logarithmic (short) path length. We show that the required routing logic can be implemented by standard prefix matching operations in hardware (TCAM) allowing for using commodity switches without any modification. Moreover, forwarding requires only a very small number of forwarding table entries, saving costly and energy-intensive TCAM.

Motivation & Objective

• Address the scalability and cost limitations of traditional Clos and Leaf-Spine data center networks that rely on high-port-density spine switches.
• Explore flat, constant-degree network topologies as alternatives to hierarchical designs to reduce hardware complexity and cost.
• Investigate whether De Bruijn graphs can provide a scalable, flat topology with efficient, hardware-implementable routing.
• Evaluate the performance trade-offs of De Bruijn-based routing versus ECMP in Leaf-Spine and random topologies under realistic traffic patterns.
• Demonstrate that De Bruijn routing can be implemented efficiently in hardware using minimal TCAM space on commodity switches.

Proposed method

• Design a flat data center network topology using De Bruijn graphs, where top-of-rack (ToR) switches are directly interconnected with constant degree (e.g., d=2).
• Implement De Bruijn routing using prefix matching operations in TCAM, enabling hardware-level forwarding with minimal table entries (max. 28 per ToR switch for 128 switches).
• Use a software-defined networking (SDN) architecture to manage and configure the De Bruijn-based network, enabling dynamic control and monitoring.
• Leverage the mathematical structure of De Bruijn graphs to guarantee logarithmic path lengths and deterministic routing without relying on probabilistic path discovery.
• Compare the De Bruijn topology with Leaf-Spine (LS/ECMP), random graphs (Random/ECMP), and De Bruijn with ECMP (DB/ECMP) using simulation-based performance evaluation.
• Simulate TCP throughput across varying network sizes (8 to 128 ToR switches) with random inter-host connections and inter-rack traffic ratios.

Experimental results

Research questions

• RQ1 Can a De Bruijn graph-based topology achieve scalable, flat interconnection of ToR switches without requiring high-port-density spine switches?
• RQ2 How efficiently can De Bruijn routing be implemented in hardware using TCAM with minimal table entries?
• RQ3 What is the performance penalty of using a flat De Bruijn topology compared to the state-of-the-art Leaf-Spine ECMP topology in terms of TCP throughput?
• RQ4 How does De Bruijn routing compare to ECMP routing in the same De Bruijn topology in terms of throughput and routing complexity?
• RQ5 How does the performance of De Bruijn networks compare to flat random graphs under realistic inter-rack traffic conditions?

Key findings

• The De Bruijn-based topology achieves logarithmic path length with a constant degree, enabling scalable interconnection of ToR switches without hierarchical spine switches.
• De Bruijn routing can be implemented using only prefix matching in TCAM, requiring a maximum of 28 forwarding table entries per ToR switch in a 128-switch topology, significantly reducing TCAM usage.
• Throughput of the De Bruijn/DBRouting topology drops to 25% of Leaf-Spine/ECMP performance at 128 switches (5120 hosts), compared to 35% for random graphs.
• De Bruijn routing (DB/DBRouting) achieves 97% to 79% of the throughput of ECMP routing on the same De Bruijn topology, indicating that the routing algorithm itself imposes minimal overhead.
• The performance gap between De Bruijn and Leaf-Spine is primarily due to longer path lengths rather than inefficient routing, as De Bruijn routing outperforms ECMP on the same topology.
• In realistic scenarios with high intra-rack traffic (80% of cloud traffic), the performance gap between De Bruijn and Leaf-Spine networks may narrow, making the cost-performance trade-off more favorable for De Bruijn in such environments.