From Serverless to Seamless: Building the Next-Generation of Cloud Systems by Eliminating Service Integration and Resource Heterogeneity Limitations

点击翻译按钮获取中文摘要

Cloud computing plays an integral role in many aspects of today's IT-driven society. The sustained growth of cloud computing is indebted to technical advances that enhanced its simplicity and scalability. The latest in this trend is serverless computing, offering maximum simplicity and scalability. It is a fast-growing paradigm that aims to decouple infrastructure management from app development. The major unsolved limitation beyond the serverless horizon is seamlessness: the ability of cloud systems to handle different usage scenarios without user intervention. My long-term research objective is to design and build this next paradigm -- seamless cloud services. This program pursues this vision by targeting two aspects of seamlessness. Developers use various service types in the cloud, including different virtual machines, microservices, and serverless offerings. Today's cloud systems lack native support for integration of diverse service types, resulting in ad hoc and "Frankenstein" hybrid applications. I aim to resolve this through seamless integration of service types (Obj. 1), using two approaches. The first is to build new cloud resource management systems that improve data locality and enhance lifetime management for hybrid applications. The second approach involves building novel data sharing systems with a focus on transparent deployment and versatility. The second area with limited seamlessness is hardware accelerators. Building a typical CPU-centric application varies greatly from coding for GPUs, etc. Future cloud systems should seamlessly leverage heterogeneous compute infrastructure without user intervention to maximize performance, efficiency, and profit. The goal for my second objective thus is seamless utilization of heterogeneous resources (Obj. 2), which I plan to achieve using three approaches. First, I will extend the current serverless model to seamlessly support different accelerators. Next, I will enable seamless, infinite scaling of accelerators. Finally, I will develop the concept of a distributed accelerator marketplace where developers can bind accelerators of different cloud providers to their applications. Successful completion of these objectives will reduce the app development time and cost of cloud services. This will have a direct impact on Canadian businesses and benefit Canada in its trade relationships. Trades between the U.S. and Canada for cloud and data storage services were valued at USD $942M in favor of the U.S. in 2019, up from USD $625M in 2018 and USD $525M in 2017. Reducing the cost of cloud services for Canadian users can decelerate this pattern. This program will train six HQP in computer systems and cloud computing areas and will prepare them for ample career opportunities in the cloud industry, related software development roles, and academic positions in software systems. The training plan is designed to recruit diverse HQP, and to provide mentorship and training in an inclusive atmosphere.