Meteion: Fast and Efficient Serverless Workflows for Latency-Critical Interactive Applications

Abstract

Serverless workflows are becoming increasingly popular in FaaS platforms. In a serverless workflow, a set of functions coordinate with each other to implement application logic. Workflows benefit from elastic scaling, load-based billing, and modular development. At the same time, existing serverless workflow frameworks suffer from significant control plane overheads due to inter-function communication, and lack effective data plane performance isolation mechanisms to prevent resource contention, leading to degraded performance. We present Meteion, a fast and efficient serverless workflow engine for latency-critical interactive applications. Meteion decouples the control and data planes, and leverages a lightweight data plane workflow engine to enable fast in-place data plane workflow execution that is asynchronous from the control plane. Meteion’s DAG scheduler uses the latency distribution and DAG structure of the workflow to provision containers in a timely manner, ensuring that functions can seamlessly execute in the data plane. Meteion also employs a per-worker-server resource manager and dynamic resource pools to provide performance isolation across function containers without incurring overheads. We evaluate Meteion across a variety of real-time data processing and interactive web applications. In all cases, Meteion outperforms prior work, reducing the end-to-end latency by 41%–75% and the provisioned memory time by 9%–23%, making serverless workflows a high performance solution for latency-critical interactive applications.

Publication
In Submission
Zhuangzhuang Zhou
Zhuangzhuang Zhou
Ph.D. Student

My research interests include cloud computing, serverless, microservices, ML for systems and computer architecture.