Concepts

Edit This Page

Poseidon-Firmament - An alternate scheduler

Current release of Poseidon-Firmament scheduler is an alpha release.

Poseidon-Firmament scheduler is an alternate scheduler that can be deployed alongside the default Kubernetes scheduler.

Introduction

Poseidon is a service that acts as the integration glue for the Firmament scheduler with Kubernetes. Poseidon-Firmament scheduler augments the current Kubernetes scheduling capabilities. It incorporates novel flow network graph based scheduling capabilities alongside the default Kubernetes Scheduler. Firmament scheduler models workloads and clusters as flow networks and runs min-cost flow optimizations over these networks to make scheduling decisions.

It models the scheduling problem as a constraint-based optimization over a flow network graph. This is achieved by reducing scheduling to a min-cost max-flow optimization problem. The Poseidon-Firmament scheduler dynamically refines the workload placements.

Poseidon-Firmament scheduler runs alongside the default Kubernetes Scheduler as an alternate scheduler, so multiple schedulers run simultaneously.

Key Advantages

Flow graph scheduling based Poseidon-Firmament scheduler provides the following key advantages:

Poseidon-Firmament Scheduler - How it works

As part of the Kubernetes multiple schedulers support, each new pod is typically scheduled by the default scheduler. Kubernetes can be instructed to use another scheduler by specifying the name of another custom scheduler (“poseidon” in our case) in the schedulerName field of the PodSpec at the time of pod creation. In this case, the default scheduler will ignore that Pod and allow Poseidon scheduler to schedule the Pod on a relevant node.

apiVersion: v1
kind: Pod

...
spec:
	schedulerName: poseidon
Note: For details about the design of this project see the design document.

Possible Use Case Scenarios - When to use it

As mentioned earlier, Poseidon-Firmament scheduler enables an extremely high throughput scheduling environment at scale due to its bulk scheduling approach versus Kubernetes pod-at-a-time approach. In our extensive tests, we have observed substantial throughput benefits as long as resource requirements (CPU/Memory) for incoming Pods are uniform across jobs (Replicasets/Deployments/Jobs), mainly due to efficient amortization of work across jobs.

Although, Poseidon-Firmament scheduler is capable of scheduling various types of workloads, such as service, batch, etc., the following are a few use cases where it excels the most:

  1. For “Big Data/AI” jobs consisting of large number of tasks, throughput benefits are tremendous.
  2. Service or batch jobs where workload resource requirements are uniform across jobs (Replicasets/Deployments/Jobs).

Current Project Stage

Features Comparison Matrix

FeatureKubernetes Default SchedulerPoseidon-Firmament SchedulerNotes
Node Affinity/Anti-AffinityYY
Pod Affinity/Anti-Affinity - including support for pod anti-affinity symmetryYYCurrently, the default scheduler outperforms the Poseidon-Firmament scheduler pod affinity/anti-affinity functionality. We are working towards resolving this.
Taints & TolerationsYY
Baseline Scheduling capability in accordance to available compute resources (CPU & Memory) on a nodeYY**Not all Predicates & Priorities are supported at this time.
Extreme Throughput at scaleY**YBulk scheduling approach scales or increases workload placement. Substantial throughput benefits using Firmament scheduler as long as resource requirements (CPU/Memory) for incoming Pods is uniform across Replicasets/Deployments/Jobs. This is mainly due to efficient amortization of work across Replicasets/Deployments/Jobs . 1) For “Big Data/AI” jobs consisting of large no. of tasks, throughput benefits are tremendous. 2) Substantial throughput benefits also for service or batch job scenarios where workload resource requirements are uniform across Replicasets/Deployments/Jobs.
Optimal SchedulingPod-by-Pod scheduler, processes one pod at a time (may result into sub-optimal scheduling)Bulk Scheduling (Optimal scheduling)Pod-by-Pod Kubernetes default scheduler may assign tasks to a sub-optimal machine. By contrast, Firmament considers all unscheduled tasks at the same time together with their soft and hard constraints.
Colocation Interference AvoidanceNN**Planned in Poseidon-Firmament.
Priority Pre-emptionYN**Partially exists in Poseidon-Firmament versus extensive support in Kubernetes default scheduler.
Inherent Re-SchedulingNY**Poseidon-Firmament scheduler supports workload re-scheduling. In each scheduling run it considers all the pods, including running pods, and as a result can migrate or evict pods – a globally optimal scheduling environment.
Gang SchedulingNY
Support for Pre-bound Persistence Volume SchedulingYY
Support for Local Volume & Dynamic Persistence Volume Binding SchedulingYN**Planned.
High AvailabilityYN**Planned.
Real-time metrics based schedulingNY**Initially supported using Heapster (now deprecated) for placing pods using actual cluster utilization statistics rather than reservations. Plans to switch over to “metric server”.
Support for Max-Pod per nodeYYPoseidon-Firmament scheduler seamlessly co-exists with Kubernetes default scheduler.
Support for Ephemeral Storage, in addition to CPU/MemoryYY

Installation

For in-cluster installation of Poseidon, please start at the Installation instructions.

Development

For developers, please refer to the Developer Setup instructions.

Latest Throughput Performance Testing Results

Pod-by-pod schedulers, such as the Kubernetes default scheduler, typically process one pod at a time. These schedulers have the following crucial drawbacks:

  1. The scheduler commits to a pod placement early and restricts the choices for other pods that wait to be placed.
  2. There is limited opportunities for amortizing work across pods because they are considered for placement individually.

These downsides of pod-by-pod schedulers are addressed by batching or bulk scheduling in Poseidon-Firmament scheduler. Processing several pods in a batch allows the scheduler to jointly consider their placement, and thus to find the best trade-off for the whole batch instead of one pod. At the same time it amortizes work across pods resulting in much higher throughput.

Note: Please refer to the latest benchmark results for detailed throughput performance comparison test results between Poseidon-Firmament scheduler and the Kubernetes default scheduler.

Feedback