Edge/cloud resource management for time-sensitive applications

Abstract

Cloud computing has become the default architecture of choice in many domains. In particular, Function-as-a-Service (FaaS), otherwise known as serverless functions, allows clients to write and deploy application functions and microservices without the need of managing their own infrastructure. With the proliferation of Internet-of-Things (IoT), many of such functions are increasingly time-sensitive in nature, and there is a need to investigate and improve on the various latencies involved in the lifecycle of a serverless function invocation. However, commercial cloud providers such as AWS, Google and Microsoft Azure are proprietary in nature, with much of the inner workings of the cloud architecture and APIs hidden from clients. As such, this project uses vHive, an open-source framework for serverless experimentation, which is designed to resemble commercial FaaS servers, and vSwarm, a serverless benchmarking function suite to be used in conjunction with vHive. vHive nodes are set up using CloudLab, and experiments are performed to illustrate the effect of cold-start delay, and queuing delay in burst invocations. Actual function runtime latency is also measured using Zipkin, and compared against the entire end-to-end latency, to obtain insight on the performance impact that function runtime has with respect to the overall latency.