kubernetes - Mesos和Kubernetes之间的调度差异

Question

前言:我的问题与this one有关，但是我想对调度的特定方面进行更深入的研究。

经过两步过程，除了Kubernetes的调度是集中式的而且Mesos的调度是分散的这一事实之外，两个项目的调度算法之间有什么区别？

我已经使用Kubernetes半年了，而我从未在实践中使用过Mesos。我了解资源提供的概念，但是我无法在Mesos和Kubernetes调度算法之间建立比较，主要是因为我对这两种工具的实现都不了解。

Answer 1

我不确定这是否具有可比性。 Kubernetes可以作为Mesos框架运行。其调度程序描述为here。它基于对节点进行过滤和排名。
Mesos的两步调度更依赖于框架算法。

Mesos向基于DRF algorithm的框架提供了报价。也可以通过使用角色和权重确定框架的优先级。

框架根据要约决定要运行的任务。每个框架都可以实现自己的算法，以使任务与商品匹配。 This is a NP hard problem

附录引用 https://medium.com/@ArmandGrillet/comparison-of-container-schedulers-c427f4f7421

Monolithic scheduling

Monolithic schedulers are composed of a singlescheduling agent handling all the requests, they are commonly used inhigh-performance computing. A monolithic scheduler generally applies asingle-algorithm implementation for all incoming jobs thus runningdifferent scheduling logic depending on job types is difficult. ApacheHadoop YARN [55], a popular architecture for Hadoop that delegatesmany scheduling functions to per-application components, is amonolithic scheduler architecture due to the fact that the resourcerequests from application masters have to be sent to a single globalscheduler in the resource master.

Two-level scheduling

A two-levelscheduler adjusts the allocation of resources to each schedulerdynamically using a central coordinator to decide how many resourceseach sub-cluster can have, it is used in Mesos [50] and was used forHadoop-on-Demand (now replaced by YARN). With this architecture, theallocator avoids conflicts by offering a given resource to only oneframework at a time and attempts to achieve dominant resource fairnessby choosing the order and the sizes of the resources it offers. Onlyone framework is examining a resource at a time thus the concurrencycontrol is called pessimistic, a strategy that is less error-prone butslower compared to an optimistic concurrency control offering aresource to many frameworks at the same time.

Shared-state scheduling

Omega grants each scheduler full access to the entire cluster,allowing them to compete in a free-for-all manner. There is no centralresource allocator as all of the resource-allocation decisions takeplace in the schedulers. There is no central policy-enforcementengine, individual schedulers are taking decisions in this variant ofthe two-level scheme. By supporting independent schedulerimplementations and exposing the entire allocation state of theschedulers, Omega can scale to many schedulers and works withdifferent workloads with their own scheduling policies [54].