Sinfonia: Cross-tier orchestration for edge-native applications

Mahadev Satyanarayanan, Jan Harkes, Jim Blakley, M. Meunier, Govindarajan Mohandoss, Kiel Friedt, Arun Thulasi, Pranav Saxena, Brian BarrittFrontiers in the Internet of Things 2022 · DOI

Sinfonia: Orchestrating the Edge Without the Headache

TL;DR

Sinfonia is an open-source framework designed to solve the "discovery and association" problem for edge-native applications like AR and autonomous drones. It provides a standardized way for mobile devices to find and connect to the most "optimal" nearby cloudlet across a three-tier architecture (Cloud, Cloudlet, Device), effectively masking the complexity of diverse infrastructure providers from the developer.

What Problem Does This Paper Address?

We are entering the era of edge-native applications—software that is simultaneously bandwidth-hungry, latency-sensitive, and compute-intensive. Think of a wearable assistant that whispers the names of people you meet or a swarm of drones navigating a forest. These apps can’t run entirely on a battery-powered device (Tier-3), but the traditional centralized cloud (Tier-1) is too far away, introducing "lag" that breaks the experience.

The solution is the cloudlet (Tier-2): a small data center located at the network edge. However, a massive problem remains: How does a device find the "best" cloudlet in real-time?

"Optimal" is a slippery term. It isn't just about physical distance; it involves network congestion, available hardware (like GPUs), business relationships (which provider do I pay for?), and even transient availability (a "just-in-time" cloudlet on a nearby truck). Currently, developers have to navigate a fragmented landscape of "walled gardens" (AWS, telcos, private clouds), making it nearly impossible to write portable edge-native code.

Methods

The authors introduce Sinfonia, a system that coordinates resources across three tiers using three main components:

  1. ST-1 (Tier-1): The "Root of Orchestration" residing in the cloud. It maintains a CloudletTable, a dynamic directory of available cloudlets, their health, and their capabilities (CPU, GPU, load).
  2. ST-2 (Tier-2): An agent running on the cloudlet. It periodically sends "keepalive" heartbeats to ST-1 to update its status in the CloudletTable.
  3. ST-3 (Tier-3): An agent on the user's device (Linux or Android).

The Workflow: When an app starts, it asks ST-3 for a backend. ST-3 queries ST-1, providing the device's location and needs. ST-1 returns a "short list" (usually 3 or fewer) of plausible candidates. Crucially, the final choice is left to the application. The app can perform its own runtime tests—like a quick ping or a small data transfer—to decide which cloudlet actually performs best at that exact moment.

Datasets

As this is a systems and architecture paper rather than a machine learning study, it does not use a training dataset. Instead, it references the OpenRTiST benchmark, an existing tool used to measure "motion-to-photon" latency in augmented reality, to justify the performance requirements of edge-native apps. The authors have released the source code for Sinfonia on GitHub for the community to evaluate.

Key Insights and Ideas

  • The Printer Driver Analogy: The authors argue that Sinfonia should be to the edge what a printer driver is to an OS. A developer should write code once, and Sinfonia should handle the "handshake" with whatever proprietary cloudlet infrastructure (the "printer") is nearby.
  • Network vs. Physical Proximity: Physical distance is a poor proxy for performance. A cloudlet 200km away might be "closer" in network terms than one 2km away if the latter is behind a congested 5G tower. Sinfonia’s two-step selection process (CloudletTable + Device-side testing) accounts for this.
  • Just-in-Time (JIT) Cloudlets: Sinfonia supports cloudlets that appear and disappear, such as a drone operator bringing a portable server to a work site. These can be discovered via local protocols like Zeroconf/Bonjour without even talking to the Tier-1 cloud.
  • End-to-End Control: A core philosophy of Sinfonia is that the app developer, not the network provider, should have ultimate control over orchestration logic.

Key Results

  • Architectural Validation: The paper demonstrates a working cross-tier workflow that successfully masks the complexity of Kubernetes-based cloudlets from the end-user application.
  • Scalability Estimates: The authors estimate a future need for $10^5$ to $10^6$ cloudlets worldwide. They argue that Sinfonia’s decentralized approach—where ST-1 roots are vendor-specific but share a common protocol—is the only way to scale to this level.
  • Open Source Release: The primary "result" is the functional, open-source implementation of the ST-1, ST-2, and ST-3 components, providing a foundation for future edge computing research.

Limitations and Future Work

  • Provisioning: The current release focuses on discovery. It assumes the backend software is already on the cloudlet or can be easily pulled. Future work is needed to handle "Just-in-Time Provisioning," where a unique "overlay" of the app is rapidly synthesized on the cloudlet at the moment of connection.
  • Privacy: Offloading data to a third-party cloudlet expands the "privacy perimeter." The authors suggest "denaturing" data (e.g., blurring faces in video) before it leaves the device, but this remains a complex challenge.
  • Trust and Micropayments: In a "free market" of cloudlets, how do you trust a random provider? The authors point toward blockchain-based micropayments and attestation-based validation as necessary future layers for the Sinfonia ecosystem.

Summary updated

Abstract

The convergence of 5G wireless networks and edge computing enables new edge-native applications that are simultaneously bandwidth-hungry, latency-sensitive, and compute-intensive. Examples include deeply immersive augmented reality, wearable cognitive assistance, privacy-preserving video analytics, edge-triggered serendipity, and autonomous swarms of featherweight drones. Such edge-native applications require network-aware and load-aware orchestration of resources across the cloud (Tier-1), cloudlets (Tier-2), and device (Tier-3). This paper describes the architecture of Sinfonia, an open-source system for such cross-tier orchestration. Key attributes of Sinfonia include: support for multiple vendor-specific Tier-1 roots of orchestration, providing end-to-end runtime control that spans technical and non-technical criteria; use of third-party Kubernetes clusters as cloudlets, with unified treatment of telco-managed, hyperconverged, and just-in-time variants of cloudlets; masking of orchestration complexity from applications, thus lowering the barrier to creation of new edge-native applications. We describe an initial release of Sinfonia ( https://github.com/cmusatyalab/sinfonia ), and share our thoughts on evolving it in the future.
BibTeX
@inproceedings{satyanarayanan2022sinfonia,
  author = {Mahadev Satyanarayanan and Jan Harkes and Jim Blakley and M. Meunier and Govindarajan Mohandoss and Kiel Friedt and Arun Thulasi and Pranav Saxena and Brian Barritt},
  title = {{Sinfonia: Cross-tier orchestration for edge-native applications}},
  booktitle = {Frontiers in the Internet of Things},
  year = {2022},
  doi = {10.3389/friot.2022.1025247},
}
Figure