Cyborg Agent Proposal

Cyborg Agent Proposal

https://blueprints.launchpad.net/openstack-cyborg/+spec/cyborg-agent

This spec proposes the responsibilities and initial design of the Cyborg Agent.

Problem description

Cyborg requires an agent on the compute hosts to manage the several responsibilities, including locating accelerators, monitoring their status, and orchestrating driver operations.

Use Cases

Use of accelerators attached to virtual machine instances in OpenStack

Proposed change

Cyborg Agent resides on various compute hosts and monitors them for accelerators. On it’s first run Cyborg Agent will run the detect accelerator functions of all it’s installed drivers. The resulting list of accelerators available on the host will be reported to the conductor where it will be stored into the database and listed during API requests. By default accelerators will be inserted into the database in a inactive state. It will be up to the operators to manually set an accelerator to ‘ready’ at which point cyborg agent will be responsible for calling the drivers install function and ensuring that the accelerator is ready for use.

In order to mirror the current Nova model of using the placement API each Agent will send updates on it’s resources directly to the placement API endpoint as well as to the conductor for usage aggregation. This should keep placement API up to date on accelerators and their usage.

Alternatives

There are lots of alternate ways to lay out the communication between the Agent and the API endpoint or the driver. Almost all of them involving exactly where we draw the line between the driver, Conductor , and Agent. I’ve written my proposal with the goal of having the Agent act mostly as a monitoring tool, reporting to the cloud operator or other Cyborg components to take action. A more active role for Cyborg Agent is possible but either requires significant synchronization with the Conductor or potentially steps on the toes of operators.

Data model impact

Cyborg Agent will create new entries in the database for accelerators it detects it will also update those entries with the current status of the accelerator at a high level. More temporary data like the current usage of a given accelerator will be broadcast via a message passing system and won’t be stored.

Cyborg Agent will retain a local cache of this data with the goal of not losing accelerator state on system interruption or loss of connection.

REST API impact

TODO once we firm up who’s responsible for what.

Security impact

Monitoring capability might be useful to an attacker, but without root this is a fairly minor concern.

Notifications impact

Notifying users that their accelerators are ready?

Other end user impact

Interaction details around adding/removing/setting up accelerators details TBD.

Performance Impact

Agent heartbeat for updated accelerator performance stats might make scaling to many accelerator hosts a challenge for the Cyborg endpoint and database. Perhaps we should consider doing an active ‘load census’ before scheduling instances? But that just moves the problem from constant load to issues with a bootstorm.

Other deployer impact

By not placing the drivers with the Agent we keep the deployment footprint pretty small. We do add development complexity and security concerns sending them over the wire though.

Developer impact

TBD

Implementation

Assignee(s)

Primary assignee:
<jkilpatr>
Other contributors:
<launchpad-id or None>

Work Items

  • Agent implementation

Dependencies

  • Cyborg Driver Spec
  • Cyborg API Spec
  • Cyborg Conductor Spec

Testing

CI infrastructure with a set of accelerators, drivers, and hardware will be required for testing the Agent installation and operation regularly.

Documentation Impact

Little to none. Perhaps on an on compute config file that may need to be documented. But I think it’s best to avoid local configuration where possible.

References

Other Cyborg Specs

History

Revisions
Release Description
Pike Introduced
Creative Commons Attribution 3.0 License

Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.