Towards All-Optical Circuit-Switched Datacenter Network Architectures with Low Energy and High Performance

Since the genesis of the “cloud”, network infrastructure has become ubiquitous across the globe and is expected to support highly customized, fine-grained applications (e.g., HPC, distributed ML, DNN, etc.) with stringent performance requirements. However, as we move to the “post-Moore’s law era” of networking, CMOS-based electrical packet-switch ASICs are struggling to cope up with the increasing capacity while maintaining low power consumption and cost. Moreover, environmental awareness makes green and long-term sustainable cloud infrastructure design an absolute necessity. This energy-critical situation has led to several recent proposals regarding all-optical circuit-switched network core design for sustainable future-generation clouds. Optical circuit-switching (OCS) technologies are the key components that make those proposals fundamentally promising, as OCS is inherently eco-friendly along with the unique advantages: a) agnostic to data rate, b) negligible/zero power consumption, c) negligible forwarding latency, and d) no need for a frequent upgrade. However, the existing OCS core-based cloud architectures pose several challenges such as a) lack of native multicast capability, b) inability to handle traffic skewness, and c) terrible tail performance of individual flows. Fundamentally, these challenges are inherent to the OCS properties and operational abstraction of existing OCS cores. First, OCS can only provide point-to-point circuits and hence cannot have multicast support. Second, round-robin OCS core architectures lack the freedom of path diversity, leading to poor performance under skewed traffic. Third, the flows incur subsequent disruption due to periodic OCS downtime, leading to unpredictable tail performance. In my thesis, I envision a holistic low-energy and high-performance cloud architecture, capable of addressing all three challenges. To address the first challenge, I propose Shufflecast: a separate optical core to support energy-efficient high-performance multicast, complementing the existing unicast-capable all-optical core. Shufflecast leverages small fanout, inexpensive, passive optical splitters to connect the Top-of-rack (ToR) switch ports, ensuring data-rate agnostic, low-power, physical-layer multicast. To address the second challenge, I propose OSSV: a combination of OCS-based core (between ToR switches) and OCS-based reconfigurable edge (between servers and ToR switches). While the OCS core is traffic agnostic and realizes reconfigurably non-blocking ToR-level connectivity, the OSSV edge reconfigures itself to rectify the incoming traffic skewness. Such spatial flexibility to reorganize the flows can largely compensate for the lack of core-level path diversity. To address the third challenge, I propose Phoenix: a more flexible OCS core and OCS edge-based architecture with precise space and time-domain control. Apart from the suitable locations, Phoenix edge can also find opportunistic moments for the flow reorganization that can minimize the OCS downtime-induced disruption. Overall, the highly flexible optical edge with both space and time-domain flexibility can significantly improve the tail performance of individual flows under realistic workloads. We extensively evaluate several aspects of these architectures with large-scale simulations and testbed implementation. We believe such holistic system design can make all-optical circuit-switched network cores widely acceptable and adoptable to the community.

Name thesis:degree_name

Doctor of Philosophy

Level thesis:degree_level

Doctoral

Discipline thesis:degree_discipline

Engineering

Grantor

Rice University

Year dc:date.issued

2024

Author dc:creator

• Das, Sushovan

Advisor dc:contributor.advisor

• Ng, T. S. Eugene

• All-optical
• Datacenter Network
• Architecture

Statement dc:rights

• Copyright is held by the author, unless otherwise indicated. Permission to reuse, publish, or reproduce the work beyond the bounds of fair use or other exemptions to copyright law must be obtained from the copyright holder.

Language dc:language.iso

eng

Handle dc:identifier.uri

https://hdl.handle.net/1911/118188

OAI identifier oai:identifier

oai:repository.rice.edu:1911/118188