High Level Goal

Traditionally, people operating in Edge Environments have had to depend on little or no supporting automation infrastructure.  The following are characteristic of Edge Environments:

• Resource constrained edge nodes
• Limited resources (CPU, battery and memory) on mobile nodes
• Limited or no end-to-end network connectivity
• Implicit assumption of WAN connectivity is not always valid
• Characterized by DIL (Disconnected, Intermittent, Limited/Low-Bandwidth)
• High cognitive load
• Application latency and fidelity become important in situations of high stress
• Bounded elasticity
• Upper bound on number of consumers known in advance
• Dynamic environment
• Static deployment topologies cannot be assumed
• Survivability is essential

Automation solutions for edge environments must operate effectively in spite of these characteristics.  Our research group has developed four primary lines of research to address the challenges of automation at the edge.  These include:

• Cyber foraging - offload of compute intensive tasks to proximate compute nodes
• CMU Cyber Foraging Research Group
• SEI Cyber Foraging Research
• SEI Tech Note on Cloudlets and the Tactical Cloud
• Collaboration with Mahadev Satyanraraynan and others at CMU
• Group Autonomy for Mobile Systems (GAMS)
• Madara project at Google Code
• GAMS project at Google Code
• SEI autonomy research
• Information Superiority to the Edge - leveraging individual and group context for information delivery
• SEI Context Aware Computing
• Collaboration with Anind Dey and others at CMU
• Edge Analytics - Compute intensive applications for edge environments supporting cyber foraging computation offload
• SEI user configured SA mashups
• Edge Analytics - real time analysis of social media

The overall goal of the current work is to develop an architecture that will enable the integration of edge automation solutions that can adapt to user needs and the changing resources available (network, computation, battery, storage, sensor capabilities, etc.).  A simplistic view of past and current edge automation is:

Current edge users have very limited connectivity.  Traditionally voice radios have been the only way of communicating, and in many cases are still the primary means of relaying information.  Even in situations where there are deployed sensors, the data from the sensors and sensor tasking is highly centralized and enabled by dedicated links from each sensor back to the supporting infrastructure.

Our approach is to break this traditional two-tier architecture into more levels to enable the forward staging of compute nodes and communications capabilities closer to the edge.  Additional wireless protocols are being explored (e.g. Delay Tolerant Networking, DTN, Nack-Based UDP, etc.) to enable much higher bandwidth wireless data communications when nodes are within range, and robust operation when they are disconnected.  Additionally, edge users should be able to directly task nearby sensors and access their data without needing to rely on distant infrastructure or cloud services.  By adding proximate compute nodes (e.g. cloudlets) very close to the edge, traditional infrastructure services can be provided within close proximity to edge users and sensors leveraging existing wireless technology and protocols.

In the next few posts I'll describe an alternative architecture and begin to detail the requirements for dynamically provisioning services to satisfy the needs of edge users all while trying to balance the inherent resource constraints involved in the scenarios we aim to support.