Towards a fully disaggregated and programmable data center

Abstract

Today, we are seeing two trends in the data center. On the one hand, applications are becoming more fine-grained, driven by the recent trend of serverless computing and microservices. On the other hand, data-center hardware is becoming more heterogeneous and customized to different computing needs. Because of these trends and for better manageability, several major data centers are moving towards a disaggregated architecture, where different hardware resources like storage and accelerators are organized as independent, network-attached pools. However, data centers today are still server-centric and relies heavily on traditional CPU-based servers. In this paper, we take a step further and explore the possibility of building a fully disaggregated data center, where every type of resource is disaggregated. Moreover, we explore the requirements and implications of making each of the disaggregated device programmable. We present guidelines and initial solutions for data center designers to navigate design trade-offs. Specifically, we decompose the overarching problem into four sub-problems and propose solutions to each of them. At the top layer, we explore two types of abstractions and propose a disaggregation-native design methodology. At the bottom layer, we describe the hardware and key features required to build disaggregated devices as well as the networking infrastructure to connect them. To bridge these two layers, we propose a static-time component that compiles different user programs into heterogeneous disaggregated devices through a disaggregation-native intermediate representation. We also propose a run-time system that manages hardware resources and schedules compiler generated execution units. We hope our proposal can pave the way for future disaggregated and programmable data center deployment.