Auction encodes Diego's long-running-process distribution algorithm
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
auctionrunner
auctiontypes
simulation
.gitignore
LICENSE
NOTICE
README.md

README.md

Auction

Note: This repository should be imported as code.cloudfoundry.org/auction.

####Learn more about Diego and its components at diego-design-notes

The auction package in this repository encodes the details behind Diego's scheduling mechanism. There are two components in Diego that participate in auctions:

  • The Auctioneer is responsible for holding auctions whenever a Task or LongRunningProcess needs to be scheduled.
  • The Rep represents a Diego Cell in the auction by making bids and, if picked as the winner, running the Task or LongRunningProcess.

The Auctioneers run on the Diego "Brain" nodes, and there is only ever one active Auctioneer at a time (determined by acquiring a lock in Consul). There is one Rep running on every Diego Cell.

The Auctioneer communicates with Reps on all Cells when holding an auction.

The Auction Runner

The auctionrunner package provides an ifrit process runner which consumes an incoming stream of requested auction work, batches it up, communicates with the Cell reps, picks winners, and then instructs the Cells to perform the work.

The Simulation

The simulation package contains a Ginkgo test suite that describes a number of scheduling scenarios. The simulation generates comprehensive output to the command line, and an SVG describing, visually, the results of the simulation run.

The simulation spins up a number of in-process SimulationReps. They implement the Rep client interface. This in-process communication mode allows us to isolate the algorithmic details from the communication details. It allows us to iterate on the scoring math and scheduling details quickly and efficiently.

Running on Diego

Instead of running the simulations by running ginkgo locally, you can run the Diego scheduling simulations on a Diego deployment itself! See the Diego Cluster Simulations repository.