Container Service for Kubernetes:Custom parameters of kube-scheduler

Prerequisites

An ACK Pro cluster, ACK Edge Pro cluster, ACK Lingjun cluster, or ACK Serverless Pro cluster that runs Kubernetes 1.20 or later is created. For more information about how to update a cluster, see Manually update ACK clusters.

Limits

Procedure

1. Log on to the ACK console. In the left-side navigation pane, click Clusters.

2. On the Clusters page, find the cluster that you want to manage and click its name. In the left-side navigation pane, choose Operations > Add-ons.

3. On the Core Components tab, find Kube Scheduler, click Configuration in the lower-right corner of the Kube Scheduler card, and then configure the parameters in the Kube Scheduler Parameters dialog box as prompted.

   The features and custom parameters vary based on the kube-scheduler version. For more information about the features supported by different kube-scheduler versions, see kube-scheduler. The following table describes the custom parameters.

   Parameter	Description	Type	Valid value	Default value
   podMaxBackoffSeconds	The maximum backoff delay for rescheduling pods in seconds. After a pod fails to be scheduled, the pod cannot be rescheduled until the delay ends.	INT	[1,100000]	10
   Preferably Use Bin Packing During Pod Scheduling	Specifies whether to schedule pods based on the binpack algorithm. For more information, see sample configurations in Custom binpack parameters.	BOOL	false true	false
   binpackPluginWeight	The weight of the node score calculated by the binpack plug-in. You must also select Preferably Use Bin Packing During Pod Scheduling.	INT	[1,100000]	100
   binpackResourceWeight	The weight of each resource when the binpack plug-in calculates the node score. You must also select Preferably Use Bin Packing During Pod Scheduling.	resourceName: STRING resourceWeight: INT	Schema validation is performed on the resourceName parameter. The value of the resourceName parameter can contain only letters, digits, periods (.), forward slashes (/), and hyphens (-). The value of the resourceWeight parameter must be an integer from 1 to 10000.	cpu:1 memory:1
   scorePluginWeights	The weight of each plug-in that is used to calculate the node score. Note The weight of the NodeResourceFit plug-in conflicts with that of the binpack plug-in. Therefore, if you select Preferably Use Bin Packing During Pod Scheduling (Pods Are Evenly Distributed among Nodes When Unselected), do not add the weight of the NodeResourceFit plug-in for this parameter.	plugin: STRING weight: INT	You can select only plug-ins that are displayed in the plugin drop-down list. The value of the weight parameter must be an integer from 1 to 10000.	plugin: NodeAffinity weight: 100
   connectionQPS	The maximum queries per second (QPS) of the Kubernetes API server.	INT	[1,1000]	50
   connectionBurst	The burst QPS when the maximum QPS of the Kubernetes API server is reached.	INT	[1,1000]	100
   percentageOfNodesToScore	The percentage of nodes suitable for pod scheduling. Default value: 0. When the default value is used, a proportion of nodes in the cluster are scored. The proportion ranges from 5% to 50% based on the cluster size.	INT	[0,100]	0
   Enable Node Scoring for GPU Sharing	Specifies whether to schedule GPU-accelerated pods to GPUs that request more memory and computing power when you use GPU sharing. To use GPU sharing, you must install the cloud-native AI suite.	BOOL	false true	true
   Enable Load-aware Scheduling (Unselected by Default)	Specifies whether to enable load-aware scheduling. To enable load-aware scheduling, you must install the ack-koordinator component.	bool	false true	false
   loadAwareThreshold	This parameter specifies the threshold for node filtering. This parameter is available after you select Enable Load-aware Scheduling.	The value consists of the resourceName and resourceWeight fields.	resourceName: Valid values are cpu and memory. threshold: Valid values range from 0 to 100. By default, this parameter is left empty, which disables node filtering.	resourceName: cpu threshold: 80
   loadAwareResourceWeight	This parameter specifies the resource weight used to calculate the node score for node sorting. This parameter is available after you select Enable Load-aware Scheduling.	The value consists of the resourceName and resourceWeight fields.	resourceName: The schema of the resourceName parameter is verified. Values can only be cpu or memory. resourceWeight: Valid values are integers ranging from 1 to 100.	cpu=1 memory=1
   loadAwareAggregatedUsageAggregationType	This parameter specifies the type of data aggregation for the statistics. Valid values: avg: calculates the average value. p50: calculates 50% of the statistics. p90, p95, and p99: calculates 90% of the statistics, calculates 95% of the statistics, and calculates 99% of the statistics.	ENUM	avg p50 p90 p95 p99	avg
   featureGates	The feature gates that are enabled by the scheduler. For more information about feature gates supported by different versions of scheduler, see the kube-scheduler documentation.	string	N/A	ACK uses the same feature gates as open source Kubernetes.

The following section provides an example on how to configure custom parameters.

Custom binpack parameters

Comparison between the binpack algorithm and spread algorithm

Configure custom binpack parameters

You can select Preferably Use Bin Packing During Pod Scheduling and specify the weight of the node score calculated by the binpack plug-in. A larger weight indicates a higher probability that pods are scheduled to the same node. You can also specify the resource names based on which the binpack plug-in calculates the node score and the weight of each resource in the node score calculation. A greater value indicates a greater influence on the node score.

On the Core Components tab of the Add-ons page, find Kube Scheduler, click Configuration, and then configure the binpack parameters in the dialog box as prompted.

If you select Preferably Use Bin Packing During Pod Scheduling without configuring the resourceName and resourceWeight parameters, the default CPU and memory settings in the following figure are used in node score calculation.

References

We recommend that you enable the load-aware scheduling, which schedules nodes based on actual resource usage rather than the amount of resources requested. For more information, see Use load-aware scheduling.