默認調度和自定義調度詳解
默認調度
默認調度是 Kubernetes 中的內置機制，它使用調度器組件來管理分配容器的節點。調度器依據以下原則選擇合適的節點：
- 資源需求 ：調度器會為每個 Pod 根據其 CPU 和內存需求選擇一個具有足夠資源的節點。
- 親和性和容忍性 ：通過親和性規則和容忍性設置，可以將 Pod 調度到滿足特定標簽或其他 Pod 運行位置要求的節點上。
- 資源分配 ：調度器會考慮節點上已運行的其他 Pod 和 Pod 的資源需求，以合理分配資源，不超出節點資源限制。
- Pod 爭用優先級 ：優先級高的 Pod 更容易被調度到可用節點上。
- 本地性 ：調度器更傾向于將新 Pod 調度到已經運行相關服務的節點上，以減少跨節點通信的開銷。
自定義調度
除了默認調度之外，Kubernetes 還允許用戶進行自定義調度，以滿足特定的業務需求。用戶可以通過以下方式實現自定義調度：
- 調度器擴展點 ：Kubernetes 支持調度器插件的機制，用戶可以使用自定義的調度器插件替換默認調度器，并實現自己的調度算法。
- 親和性和容忍性調度規則 ：通過在 Pod 的調度規范中指定親和性和容忍性規則，用戶可以要求 Pod 被調度到特定的節點上。
- 節點選擇器 ：通過使用節點選擇器標簽，用戶可以將一組特定的節點應用于特定類型的工作負載。
調度原理
Kubernetes 的調度器基于優先級和可行性評分算法來決定將 Pod 調度到哪個節點。調度流程如下：
- 預選階段（Predicates） ：調度器對每個節點應用預選謂詞函數，檢查是否滿足 Pod 的資源需求、親和性規則等。不滿足條件的節點將被標記為不可調度。
- 優選階段（Priorities） ：調度器對所有可調度的節點應用優選函數，為每個節點分配優先級。優選函數根據爭用優先級和伸縮策略等標準為節點打分。
- 選定階段（Binding） ：調度器選擇優先級最高的節點將 Pod 調度到該節點。如果存在多個具有相同最高優先級的節點，則將應用綁定函數來決定最終調度。
On the other hand
Kubernetes is a powerful container orchestration platform that automates the deployment and management of containerized applications. An essential aspect of Kubernetes is its scheduling capability, which determines where to place containers within the cluster. In this blog, we will explore the default scheduling mechanism in Kubernetes and also understand how custom scheduling can be leveraged to meet specific business requirements.
Default Scheduling
Default scheduling is the built-in mechanism in Kubernetes that utilizes the scheduler component to manage container placement on nodes. The default scheduler follows these principles to select suitable nodes for pods:
- Resource Requirements : The scheduler matches each pod’s CPU and memory requirements with nodes having sufficient resources.
- Affinity and Tolerations : By using affinity rules and tolerations, pods can be scheduled on nodes with specific labels or meet certain requirements defined by other pods.
- Resource Allocation : The scheduler takes into account the already running pods and their resource demands on each node, ensuring fair distribution without exceeding node resource limits.
- Pod Preemption Priority : Pods with higher priority have a better chance of being scheduled on available nodes.
- Locality : The scheduler prefers to place new pods on nodes already running related services to minimize inter-node communication overhead.
Custom Scheduling
While default scheduling covers most scenarios, Kubernetes allows users to implement custom scheduling to cater to specific business needs. Custom scheduling can be achieved through the following approaches:
- Scheduler Plugins : Kubernetes supports scheduler plugin mechanisms that enable users to replace the default scheduler with their own implementation, incorporating custom scheduling algorithms.
- Affinity and Tolerations Scheduling Rules : By specifying affinity and tolerations rules in pod scheduling specifications, users can enforce pods to be scheduled on specific nodes.
- Node Selectors : By using node selector labels, users can apply a specific set of nodes for specific types of workloads.
Scheduling Principles
The Kubernetes scheduler utilizes priority and feasibility scoring algorithms to determine which node a pod should be scheduled on. The scheduling process includes the following stages:
- Predicates Stage : The scheduler applies predicate functions to each node to evaluate if pods meet resource requirements, affinity rules, and other conditions. Nodes that fail the predicates are marked as not scheduling candidates.
- Priorities Stage : The scheduler assigns priority scores to all nodes based on various criteria such as contention, autoscaling policies, etc.
- Binding Stage : The scheduler selects the node with the highest priority score to bind the pod. In case of multiple nodes with the same highest priority, binding functions come into play for the final decision.
)
