Year: 2025
Type: dissertation
Abstract: Modern data center applications are too large to fit on a single device [1]. However, deploying an application using multiple devices requires additional design work to ensure that the devices can communicate and collaborate effectively. To ease this deployment process, the Laniakea framework is designed to deploy large applications across a heterogeneous mix of network-connected Central Processing Units (CPUs) and Field Programmable Gated Arrays (FPGAs) [2]. Laniakea automatically adds the interconnect and orchestration logic required to maintain application functionality, and supports multiple compilation flows targeting a variety of hardware devices. To support larger application deployments, the framework proposes several features to decouple application size from hardware constraints, such as executing multiple application components sequentially using the same hardware. However, these features require Laniakea to have a sophisticated control plane, capable of making deployment decisions and configuring hardware at runtime. In this thesis, we enable this increased level of management through the creation of a communications protocol designed for Laniakea's control plane. We design the Laniakea Control Protocol, a control plane communications protocol that supports intra- and inter-cluster request/response style messaging. To ensure that LCP messages arrive at their destination, we develop the Reliability Protocol-Node (RPN), which provides exactly-once LCP message delivery guarantees over unreliable networks. Finally, we discuss how these protocols can be used to enable Laniakea's proposed features when they are implemented.
Source: TSpace (University of Toronto)
Author: Hai, Kin Hang Justin
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7132914925