How Koordinator Supercharges ACK Container Scheduling and Resource Efficiency

What is Koordinator

Koordinator is an open‑source container scheduling framework originated from Alibaba’s extensive experience in mixed‑workload placement. It aims to improve performance of latency‑sensitive services and batch jobs while reducing overall cluster resource consumption.

Background

Research on container scheduling started at Alibaba in 2011, and a dedicated mixed‑workload scheduling project began in 2016. After multiple iterations the technology matured into a cloud‑native architecture that now runs on clusters with more than ten million CPU cores, achieving average CPU utilization above 50 % during peak events such as Double‑11.

Integration with Alibaba Cloud ACK

In April 2022 the project was open‑sourced as “Koordinator” and integrated into Alibaba Cloud Container Service for Kubernetes (ACK). The integration provides a low‑cost entry point for mixed‑workload scenarios and is used in production by many enterprises.

Architecture and Core Components

Koordinator Manager (deployed as a Deployment) – central control plane that runs two controllers:

SLO Controller : manages resource oversubscription per node and enforces differentiated SLO policies.

Recommender : collects historical usage, builds resource‑profile models and generates request/limit recommendations.

Koordinator Descheduler (Deployment) – performs pod‑level rescheduling based on configurable policies.

Koordlet (DaemonSet) – runs on each node to enforce oversubscription, fine‑grained scheduling decisions and QoS guarantees.

Key Technical Features

QoS‑Aware Scheduling & Rescheduling

Load‑aware scheduling : during the scoring phase the scheduler evaluates real‑time node metrics (CPU, memory, disk, network) and prefers nodes with lower observed load, preventing hot‑spots.

Deterministic rescheduler : based on user‑defined policies (time windows, priority, disruption budgets) it safely migrates pods to achieve a balanced resource distribution.

Resource Profiling

The profiling pipeline continuously collects container usage metrics, aggregates them over configurable windows (e.g., 24 h), and applies statistical models to estimate peak demand. The resulting profile is used to suggest optimal requests and limits, reducing over‑provisioning and improving cluster utilization.

Differentiated SLO Mixing

Resource‑tiered scheduling : workloads are classified into tiers (e.g., latency‑critical, best‑effort) and receive secondary resource allocations based on their real‑time profiles.

Isolation & interference suppression : the system enforces CPU, memory, disk and network isolation for each tier, limiting the impact of noisy neighbors.

Release v1.1.1‑ack.1 Highlights

This version adds native support for load‑aware scheduling and the deterministic rescheduler, and integrates the resource‑profiling suite into the ACK cost‑optimization toolkit.

Typical Usage Flow

Deploy the three components (Manager, Descheduler, Koordlet) via standard Kubernetes manifests or Helm charts.

Enable the SLO Controller and Recommender in the Manager spec.

Allow Koordlet to collect per‑node metrics (CPU burst, memory usage, I/O, network).

Periodically run the Recommender; it writes recommendation CRDs that can be applied with kubectl apply -f.

Configure the Descheduler policies (e.g., PodLifeTime, NodeUtilizationThreshold) to trigger automatic pod migrations.

Repository and Documentation

Source code and release artifacts are hosted at: https://github.com/koordinator-sh/koordinator Official project site with detailed documentation:

https://koordinator.sh