October 5, 2025

FaaSNet: Scalable and Fast Provisioning of Custom Serverless Container Runtimes at Alibaba Cloud Function Compute

Decentralize container provisioning across host VMs in function-tree structures to eliminate cold-start bottleneck

Venue: ATC 2021

Topic: Large container image pulls dominate FaaS cold-start costs. Existing P2P approaches require dedicated root servers that don’t fit FaaS infrastructure. FaaSNet decentralizes provisioning across VMs organized in function-based tree structures.

Summary

Pulling large container images from a backing store causes significant cold-start latency (up to several minutes under contention). Existing P2P solutions require powerful, centralized root servers — incompatible with FaaS’s dynamic pool of resource-constrained VMs. FaaSNet: a lightweight, adaptive middleware that decentralizes provisioning across VMs in function trees (FTs), with a tree-balancing algorithm to handle dynamic VM membership.

Background

Cold start and container image pull

• Container image pull cost dominates most functions’ cold-start overhead.
• Large images amplify the problem significantly.

Existing approaches and their limits

• Kraken, DADI, Dragonfly: P2P approaches for accelerating container provisioning at scale.
  • Require one or more dedicated, powerful root servers for data seeding, metadata management, and coordination.
  • Direct application to FaaS is problematic:
    1. Increases total cost of ownership (TCO) and introduces a performance bottleneck.
    2. FaaS uses a dynamic pool of resource-constrained VMs — VMs join and leave at any time. Existing P2P solutions lack the required adaptivity.

Key Idea

Function Tree (FT)

• A logical, tree-based network overlay consisting of host VMs.
• Container provisioning is decentralized across all VMs in the tree in a scalable manner.

Tree balancing algorithm

• Dynamically adapts FT topology to accommodate VM joining and leaving.
• Ensures high adaptivity without a centralized root server.

Design

FaaSNet architecture

1. Function Tree (FT): organizes host VMs into a tree overlay per function type.
   • Each node can receive and forward container images to its children.
   • No single point of failure or bottleneck.
2. Tree balancing: on VM join/leave, the algorithm restructures the FT to maintain balance and provisioning efficiency.
3. Lightweight middleware: integrates with existing FaaS infrastructure without requiring extra dedicated hardware.

Evaluation

• Deployed and evaluated at Alibaba Cloud Function Compute.
• Significantly reduces cold-start latency for large container images compared to pulling from a centralized backing store.
• Scales efficiently as the number of concurrent function invocations grows.

Meeting Notes

(to be filled)