Reinventing Sustainable Data Center Networks for the AI Age
The rapid growth of artificial intelligence (AI) workloads is driving the need for highly scalable, low-latency, low-complexity, resilient and sustainable data center networks (DCNs). AI training and inference environments is placing significant pressure on IP-based routing architectures designed for technologies and communications of decades ago. While substantial efforts focus on expanding AI adoption across domains, parallel initiatives aimed at mitigating the AI-driven data center energy crisis largely emphasize improved cooling technologies and increased power generation capacity. The proposed research project targets the root cause of these challenges—the continued reliance on retrofitted, heavyweight legacy protocols, which introduce inefficiencies that degrade DCN performance and contribute significantly to their escalating power consumption.
The project Multi-Root Meshed Tree Protocol (MR-MTP) offers a simple IP agnostic solution by adopting a novel technique to establish multiple trees rooted at leaf nodes in a folded-Clos topology DCN that virtually mesh at the spines to provide multiple loop-free forwarding paths between server racks. MR-MTP replaces seven protocols in the traditional IP-based DCN router protocol stack, resulting in heavy reduction in router hardware and software, leading to highly reduced operational complexity and a proportional reduction in energy consumption and equipment cost. MR-MTP’s auto-configuration reduces human errors and simplifies network maintenance, management, and troubleshooting and the IP agnostic approach mitigates vulnerabilities associated with IP-layer attacks.
AI workloads dictate bypassing traditional IP/TCP stack, minimizing software overhead, designing deterministic fabrics and collapsing multiple protocols among others. Infiniband and RoCE(Ethernet +Remote Direct Memory Access) are two technologies targeting novel approaches. While InfiniBand demonstrates the benefits of a non-IP, simplified, high-performance fabric, and RoCE retrofit Ethernet to meet AI demands, both approaches do not address the root cause—protocol complexity. MR-MTP collapses multiple protocols to a single, topology-aware mechanism, enabling simpler, faster, and more secure data center networks. Hyperscale system challenges directly expose the limitations MR-MTP is trying to eliminate. MR-MTP is the next step beyond RDMA fabrics and IP-based DCNs—where the network is no longer configured through protocols but emerges directly from topology.
C-coded MR-MTP was tested using FABRIC large-scale experimental testbed to emulate near–real-world deployment scenarios and its performance compared with the popular protocol suite Border Gateway Protocol (BGP), Equal Cost Multipath Protocol (ECMP) and Bidirectional Forwarding Detection (BFD) in folded-Clos DCN topologies. Results demonstrate the significant performance improvements and stability of MR-MTP. (Code and FABRIC test scripts available on GitHub).
MR-MTP supports interoperability with DCNs operating on conventional protocols- allowing incremental deployment. With a Technology Readiness Level (TRL) of 6 and recognition by DARPA’s Expedited Research Implementation Series (ERIS) as a Breakthrough Technological Advancement, MR-MTP is positioned for real-world deployment.
Date and Time
Location
Hosts
Registration
Add Event to Calendar
Speakers
Nirmala of Rochester Institute of Technology
Biography:
Short Bio: Dr. -Ing Nirmala Shenoy earned her Bachelor's and Master's degrees in Engineering from Madras University, India. She was a Senior Research Scientist and Assistant Director at the Central Electronics Engineering Research Institute, Council of Scientific and Industrial Research in India. She was the recipient of the prestigious Deutscher Akademischer Austausch Dienst (DAAD) fellowship and pursued Ph.D. in Computer Science with a focus on network protocols at the University of Bremen, Germany. Dr. Shenoy taught at the Information Communication Institute of Singapore, a collaboration of AT&T Bell Labs (U.S) and the National Computer Board of Singapore. She has held teaching and research positions at Australian universities before joining the Rochester Institute of Technology, Rochester, New York in 2001. She currently is a professor in the iSchool, School of Information, Golisano College of Computing and Information Sciences and heads the Lab for Networking and Security. Her research includes mobility management, network architectures and protocols for wired, wireless ad hoc and sensor networks with a focus to simplify network operations and protocols through novel techniques aligned to current day technologies and networks. She chaired the IEEE project 1910.1 “Meshed Tree Bridging for loop-avoidance in switched networks”. Her research received funding from the Department of Defense, National Science Foundation, New York State, Boeing, and Cisco among others
Email:
Address:102, Lomb Memorial drive, , Rochester, , New York, United States, 14623