Scheduling

Updated on 2025-02-28 GMT+08:00

View PDF

Basic kube-scheduler configurations and Volcano-backed advanced scheduling are available. You can enable advanced scheduling functions such as bin packing, priority-based scheduling and preemption, AI task performance enhancement, and heterogeneous resource management for improved cluster resource utilization at low costs.

Scheduler Configuration

NOTE:

Only kube-scheduler supports this configuration.

**Table 1** Parameters
Item	Parameter	Description	Value
QPS for communicating with kube-apiserver	kube-api-qps	QPS for communication with kube-apiserver	If the number of nodes in a cluster is less than 1,000, the default value is 100. If the number of nodes in a cluster is 1,000 or more, the default value is 200.
Burst for communicating with kube-apiserver	kube-api-burst	Burst for communication with kube-apiserver	If the number of nodes in a cluster is less than 1,000, the default value is 100. If the number of nodes in a cluster is 1,000 or more, the default value is 200.

Business Priority Scheduling

Scheduling based on priority

This is a basic scheduling capability and cannot be disabled. The scheduler preferentially guarantees the running of high-priority pods, and will not evict low-priority pods that are running. For details, see Priority-based Scheduling and Preemption.

Resource Utilization Optimization Scheduling (Supported by the Volcano Scheduler)

Enable boxing strategy (binpack)

With this option enabled, the cluster scheduler schedules pods to nodes that have the most requested resources. This reduces resource fragments on each node and improves the resource utilization of the cluster. For details, see Bin Packing.

For details about the bin packing weight and weights for each resource to score nodes, see Table 2.

**Table 2** Bin packing weight
Item	Description	Default Value
Binpack Scheduling Strategy Weight	A larger value indicates a higher weight of the bin packing policy in overall scheduling.	10
CPU Weight	A larger value indicates a higher cluster CPU usage.	1
Memory Weight	A larger value indicates a higher cluster memory usage.	1
Add custom resource type	Other custom resource types requested by pods, for example, nvidia.com/gpu. A larger value indicates a higher usage of the specified cluster resource.	-

AI Task Performance Enhancement Scheduling (Supported by the Volcano Scheduler)

Fair Scheduling Policy (drf)

Dominant Resource Fairness (DRF) is a scheduling algorithm based on the dominant resource of a container group. It supports fair allocation of multiple types of resources and is suitable for batch AI training and big data jobs. DRF is suitable for batch process small scale services like single AI model training and single big data computing and query, because it preferentially considers the throughput of services in clusters.

DRF helps you enhance the service throughput of clusters and improve service performance. For details, see DRF.

Workload Group Scheduling Policy (Gang)

Gang scheduling meets the scheduling requirements of "All or nothing" in the scheduling process and avoids the waste of cluster resources caused by arbitrary scheduling of pods. It is mainly used in scenarios that require multi-process collaboration, such as AI and big data scenarios.

Gang scheduling effectively resolves pain points such as resource waiting or deadlocks in distributed training jobs, thereby significantly improving the utilization of cluster resources. For details, see Gang.

Heterogeneous Resource Scheduling (Supported by the Volcano Scheduler)

Support GPU resource scheduling

To use this feature, install the CCE AI Suite (NVIDIA GPU) add-on in the cluster first. With this option enabled, GPUs can be used for AI training jobs, and the scheduler provides full GPU dispatch and GPU sharing to improve resource utilization.

Parent topic: Settings

Previous topic: Network

Next topic: Auto Scaling