Ksched: A Firmament Prototype

Firmament is a new apporach to cluster scheduling that models the problem of scheduling as a network flow optimization problem. Firmament aims on providing:

• Policy optimal scheduling
• Lower latency on scheduling decisions
• Flexibility through pluggable policies

Ksched is an experimental reimplementation of the scheduler from it's C++ implementation to Go. The goal of this project is to integrate Firmament in Kubernetes as an alternative scheduler.

Current State of Project:

The project so far is an early stage prototype with all the mechanisms for performing multiple scheduling iterations.

The scheduler layer has a minimal interface with the Kubernetes API allowing it to batch schedule pods. Currently the implementation has no support for sophisticated policies and will perform a trivial first-fit policy to assign pods to nodes in the cluster.

Trying it Out:

To get the scheduler up and running there are two ways to currently test it out.

Option 1: Running on a live cluster:

You can test the scheduler by running it inside of a container on the kubernetes master node. You can build the image from build/Dockerfile yourself or as described below use our hosted image.

On the master node pull the image.

docker pull hasbro17/ksched:v0.6

Run the container on the host network, waiting in background mode.

docker run --net="host" --name="ksched" -d hasbro17/ksched:v0.6 tail -f /dev/null

You will need to pause the pre-existing kubernetes scheduler's container before trying to run ksched.

docker pause <container-ID>

Get a shell into the ksched container.

docker exec -it ksched /bin/bash

Run the init script to clone and build the scheduler.

/root/init.sh

There should be two binaries present in the ksched project at /root/go-workspace/src/github.com/coreos/ksched

The first k8sscheduler, is the scheduler whose flags are specified in cmd/k8sscheduler/scheduler.go. Run this binary to start the scheduler.

k8sscheduler -fakeMachines=false

The scheduler should start up and wait for unscheduled pods at this point.

To generate a large number of pod requests you can use the binary podgen.

podgen -numPods=<number-of-pods> -image=nginx

Option 2: Run with Kubernetes API server:

You can test the scheduler without the real cluster by only having the kube-api binary running. The setup is a little more involved for this case.

You will need to have the same environment set up as is for the ksched image described by build/Dockerfile. Use that as a guide for your setup.

• Setup the Flowlessly solver binary in the correct location: /usr/local/bin/flowlessly/.
• Setup the Kubernetes(v1.3) source at the following location in your go workspace: $GOPATH/src/k8s.io
• Get the ksched source: go get github.com/coreos/ksched (or from the mirror repo github.com/hasbro17/ksched-mirror)
• Generate the proto files by running proto/genproto.sh