Required Readings



Tiers in Computing

Demand for lower latency drives compute towards where data is collected

Clouds -> Cloudlets (luggable, vehicular, mini-datacenter) -> IOT (drones, security cameras, smartphones) -> RFID tags

Why Edge Computing

Newer workloads (HD Video, AR/VR, SmartCity & Automation) increase demand for bandwidth-intensive and latency-sensitive systems.

Working from home also shifts connectivity needs.

Closing the Latency-Bandwidth Gap

5G helps, but might be too little, too late.

Transitioning to newer software stacks has been traditionally slow, but is getting faster due to trends of Open Source Software and commodity hardware.

Moving computation to the closer to devices at the users is called edge computing.

Is Edge Computing New

CDNs have existed for a long time. CDNs are basically just caches, and still don’t hit the scale of edge computing.

Edge Computing Drivers

  • Speed of Light
    • Data can only move so fast
  • Data growth
  • New Apps
    • 5G
    • Video
    • AR/VR
    • IoT
    • Cognitive tasks
    • Enterprise/private LTE
  • Data Sovereignty laws

Distributed Edge Computing

Edge != Cloud

Scale and geo-distribution are different

“Chatty” protocols are not appropriate

  • heartbeats, etc

Edge is not elastic

Mobility, device churn, reliability


Localization, contextualization


IoT and Distributed Transactions

Smart devices -> edge gateways -> Cloud

  • Sensors and Actuation

Physical state != application state


High level programming abstraction and model

        [sensorList, timeWindow, sensingPolicyt],

onSuccess() and onFailure() methods

Provide atomic durability of actuations

Transactuation Invariants

  • Sensing invariant
    • Transactuation executes only when staleness of its sensor reads is bounded, as per specified sensing policy.
  • Sensing policy
    • How much staleness is acceptable?
    • How many failed sensors are acceptable?
      • At least one CO2 sensor must be read within last 5 minutes
  • Actuation invariant
    • When a transactuation commits its app states, enough actuations have succeeded as per actuation policy.
  • Actuation policy
    • How many failed actuations are acceptable?
      • At least one alarm should turn on.

Evaluation of Transactuations

Evaluation Questions:

  • Impact on programmability
    • Typically lines of code
  • Impact on performance during failure-free execution
    • Benchmark before and after adding safety measures
  • Impact on correctness during failures
    • Unit/e2e tests before and after adding safety measures


ApplicationOriginal AppOriginal App + ConsistencyTransactuations
Rise and Shine7219568
Whole house fan2917626
Thermostat auto off7019868