Serverless in the Wild - Characterizing and Optimizing the Serverless Workload

Explore a comprehensive analysis of serverless computing in this 23-minute conference talk from USENIX ATC '20. Dive into the characteristics and optimization of serverless workloads at a large cloud provider, specifically Microsoft Azure. Learn about Function as a Service (FaaS) popularity, its challenges for cloud providers, and the importance of understanding FaaS workload characteristics. Discover key findings from Azure Functions' production workload, including the wide range of invocation frequencies and resource management implications. Examine a proposed practical resource management policy aimed at reducing function cold starts while optimizing resource utilization. Gain insights into serverless computing trends, workload heterogeneity, memory usage patterns, and function execution durations. Understand the limitations of fixed keep-alive policies and explore innovative approaches like the Hybrid Histogram Policy. Follow the implementation in OpenWhisk and simulation results to grasp the potential improvements in serverless resource management.

Intro
What is Serverless?
So what are people doing with Faas?
If you are a cloud provider
Stepping Back: Characterizing the Workload
Invocations per Application
Apps are highly heterogeneous
What about memory?
Function Execution Duration
Key Takeaways
Cold Starts and Resource Wastage
What do serverless providers do?
Fixed Keep-Alive Policy
Fixed Keep-Alive Won't Fit All
Fixed Keep-Alive Is Wasteful
A Histogram Policy To Learn Idle Tinil
The Hybrid Histogram Policy
More Optimal Pareto Frontier 130
Implemented in OpenWhisk
Simulation
Closing the loop