Doctoral Thesis Proposal - Shawn Chen

— 3:30pm

Location:
In Person - Gordon Bell Conference Room, Gates Hillman 5117

Speaker:
SHAWN SHUOSHUO CHEN , Ph.D. Student, Computer Science Department, Carnegie Mellon University
https://shuoshuc.github.io/

Reshaping Data Center Networks with Reconfigurability

Data center networks are fundamental to cloud computing—they tightly couple compute and storage with high bandwidth and low latency. The demand for data center network bandwidth is continuously growing, driven by the proliferation of data-intensive applications like AI/ML and video streaming. However, electrical packet switches struggle to deliver the total bandwidth required by the growing demands because a “plateauing” Moore’s Law limits I/O density and high-speed memory capacity. Moreover, the sheer scale of modern data center networks makes electrical packet-switched networks increasingly expensive and power-hungry. Reconfigurable optical switching technology is a promising alternative, offering the potential for higher bandwidth, reduced energy consumption, and runtime reconfigurability. Reconfigurable data center networks (RDCNs) combine the benefits of both optical and packet switches to accommodate diverse traffic patterns and enhance network performance. 

This thesis addresses the limitations of current network designs in RDCNs by revisiting underlying assumptions and redesigning core network components, focusing on transport, traffic engineering, and topology. First, we present Time-division TCP (TDTCP), a new transport protocol that adapts to the fluctuating bandwidth and latency in demand-oblivious RDCNs by maintaining independent network states for each time-division multiplexed path. Second, we tackle traffic engineering in demand-aware RDCNs with approaches that help implement complex traffic engineering solutions in switches with minimum precision loss. Third, we propose a flexible machine learning job scheduling mechanism for reconfigurable clusters based on torus topologies, ensuring optimal job performance while mitigating resource fragmentation. Together, these innovations aim to unlock the full potential of RDCNs, achieving higher performance, cost-efficiency, and scalability for future data center workloads. 

Thesis Committee

Srinivasan Seshan (Chair)
Peter Steenkiste
Tim Dettmers
Minlan Yu (Harvard University)

Additional Information

Add event to Google
Add event to iCal