Resource managers exist to collect computing resources from multiple disparate machines and abstract their use into a unified interface for distributed applications. Existing resources managers are focused on performing this task for cloud-based application.
This is a new resource manager that attempts to break that mold and extend resource management across networks and computing classes. While the underlying ideas of resource management are not new (and are heavily based on Apache Mesos), the implementation details are. The resource manager agent is minimal such that it can be run on ultra-small computing devices. Diverse networking protocols are supported so that one can aggregate resources across bandwidth constrained networks and behind NATs. We intend to support executors such as WASM in addition to traditional container environments so that multiple tasks can be run in memory constrained environments.
The resource manager consists of a master node, multiple agent nodes, and frameworks which facilitate the scheduling of applications (which are broken down into tasks on the) on the agent nodes.
The EdgeRM Master node collects resources from agent nodes and offers those resources to frameworks which can schedule tasks on those resources. Currently the EdgeRM Master is implemented in python and supports both HTTP and COAP networking.
If the entire deployment is within a local network, this could be deployed locally, but we expect it will be most useful if deployed on a publicly accessible server.
EdgeRM agent nodes offer resources on which tasks can be scheduled. They can range from small embedded nodes up to traditional Linux servers. Agent nodes support one or mode execution environments (currently Docker and WASM are supported), and can advertise traditional resources such as compute and memory, and less traditional resources, such as sensor and actuator devices.
Currently agent nodes are implemented in python supporting a Docker execution environment and in C on the zephyr OS supporting a WASM execution environment.
Resources are anything that needs to be shared are partitioned by tasks running on an agent node, and attributes are added information about the nodes on which these resources exist. While the EdgeRM protocol and master do no enforce the usage of specific resources and attributes, without standardization of these types there would be little portability between frameworks and agent nodes.
A current list of the resources and attributes we use in our agent and framework implementations are here.
./master - the EdgeRM master implementation written in python
./agent - multiple EdgeRM agent implementations
./proto - EdgeRM Protobuf definitions
./frameworks - several example frameworks which use EdgeRM
./edgerm - the edgerm python library code. Utilities to create frameworks and schedulers.
./test - full systems tests
./media - media for this project
./docs - more documentation about details of the project