Dirigent: Lightweight Serverless Orchestration

Overview

• FaaS (Function as a Service) platforms can quickly initialize function sandboxes, but scheduling functions in real FaaS clusters can take much longer.
• The current approach of building FaaS cluster managers on top of legacy orchestration systems like Kubernetes leads to high scheduling delays when there is a lot of sandbox creation.
• Existing cluster managers use complex abstractions and components that become a bottleneck for the frequent state changes in FaaS environments.

Plain English Explanation

When you use a FaaS service like AWS Lambda, the platform can quickly create a tiny, isolated "sandbox" environment to run your code. However, the overall process of scheduling that function to run in a real FaaS cluster can take much longer, often orders of magnitude higher than the sandbox creation time.

The researchers found that the current way of building FaaS cluster management systems, using existing orchestration platforms like Kubernetes, leads to these high scheduling delays. These general-purpose cluster managers use complex internal structures and processes to manage and keep track of the state of the entire cluster. But in a FaaS environment, the cluster state is constantly changing as new function sandboxes are quickly created and destroyed. This frequent flux in the cluster state becomes a bottleneck for the cluster manager.

Technical Explanation

The researchers analyzed the root causes of performance issues in existing FaaS cluster managers and proposed a new system called Dirigent to address them. Dirigent has three key principles:

1. Optimized Abstractions: Dirigent simplifies the internal cluster manager abstractions to streamline state management, unlike the hierarchical and complex designs of generic cluster managers.

2. Eliminating Persistent Updates: Dirigent avoids making persistent state updates on the critical path of function invocations. This is possible because FaaS abstracts away the sandboxes from users, relaxing the need for exact state reconstruction.

3. Monolithic Design: Dirigent runs a monolithic control and data plane, minimizing internal communication overhead and maximizing throughput, unlike the distributed architectures of existing platforms.

The researchers compared Dirigent to state-of-the-art FaaS platforms and found that it reduces the 99th percentile per-function scheduling latency by 2.79x compared to AWS Lambda. Dirigent can also spin up 2500 sandboxes per second at low latency, which is 1250x more than the Knative platform.

Critical Analysis

The paper provides a thorough analysis of the performance issues in current FaaS cluster management systems and proposes a thoughtful solution in Dirigent. However, the researchers acknowledge that Dirigent's design choices, such as the monolithic architecture, may have limitations in terms of scalability and fault tolerance compared to more distributed systems.

Additionally, the paper focuses on optimizing function scheduling latency, but does not address other important FaaS considerations, such as resource utilization, cost-efficiency, or support for advanced features like stateful functions. Further research may be needed to evaluate Dirigent's performance and applicability in a broader range of FaaS use cases.

Conclusion

The Dirigent system proposed in this paper presents a novel approach to FaaS cluster orchestration, addressing the performance bottlenecks of existing solutions. By simplifying internal abstractions, eliminating persistent state updates, and using a monolithic design, Dirigent achieves significant improvements in function scheduling latency and sandbox creation throughput.

These advancements in FaaS cluster management could have important implications for the adoption and performance of serverless computing, enabling more responsive and efficient execution of cloud-hosted functions. As the demand for serverless architectures and containerized microservices continues to grow, innovations like Dirigent could help accelerate microservices and optimize the scheduling of DNN workflows in FaaS environments.