Network Topologies for Composable Data Centers

• Main Authors: Opeyemi O. Ajibola, Taisir E. H. El-Gorashi, Jaafar M. H. Elmirghani
• Format: Article
• Language: English
• Published: IEEE 2021-01-01
• Series: IEEE Access
• Subjects: Composable data centers; disaggregated data centers; energy efficient networks; data center networks; MILP; optical communication
• Online Access: https://ieeexplore.ieee.org/document/9519656/

Suitable composable data center networks (DCNs) are essential to support the disaggregation of compute components in highly efficient next generation data centers (DCs). However, designing such composable DCNs can be challenging. A composable DCN that adopts a full mesh backplane between disaggregated compute components within a rack and employs dedicated interfaces on each point-to-point link is wasteful and expensive. In this paper, we propose and describe two (i.e., electrical, and electrical-optical) variants of a network for composable DC (NetCoD). NetCoD adopts a targeted design to reduce the number of transceivers required when a mesh physical backplane is deployed between disaggregated compute components in the same rack. The targeted design leverages optical communication techniques and components to achieve this with minimal or no network performance degradation. We formulate a mixed integer linear programming (MILP) model to evaluate the performance of both variants of NetCoD in rack-scale composable DCs that implement different forms of disaggregation. The electrical-optical variant of NetCoD achieves similar performance as a reference network while utilizing fewer transceivers per compute node. The targeted adoption of optical technologies by both variants of NetCoD achieves greater (4 - 5 times greater) utilization of available network throughput than the reference network which implements a generic design. Under the various forms of disaggregation considered, both variants of NetCoD achieve near-optimal compute energy efficiency in the composable DC while satisfying both compute and network constraints. This is because marginal concession of optimal compute energy efficiency is often required to achieve overall optimal energy efficiency in composable DCs.