OPTIMIZING COLD START LATENCY IN SERVERLESS COMPUTING
Abstract
Serverless computing has gained traction in public cloud offerings, including AWS, Azure, and GCP, in the past few years. Consumers of these platforms cherish the ability to write multiple functions without much need to write boilerplate code or to manage these servers. However, despite its benefits, cloud computing suffers high latency when reacting to intermittent events, due to the cost of deploying function code and data to new instances and the cost of initializing the sandboxed function runtime--known as the cold start latency. Cold start latency poses a major challenge to the burst-out performance of serverless applications. In this thesis, we analyze the components of cold start latency in commercial serverless platforms to understand the factors of networking and the impact of function binary sizes. We further propose a solution to build a framework to reduce the latency of data transfer when initializing a function on a new instance. We use the idea of deduplication and data encoding, to reuse data blocks from a global corpus of frequently used blocks, or to reuse blocks that previously occur on the same machines. We build a framework for general optimization of data movement between cloud servers and demonstrate that with a high data reuse rate, the network component of the cold start latency can be significantly reduced.
Citation
Bhat, Nikhil Premanand (2020). OPTIMIZING COLD START LATENCY IN SERVERLESS COMPUTING. Master's thesis, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /192194.