[Cloud Networking Notes] Management and sharing of network infrastructure in cloud data centers

来源:互联网 发布:linux服务器监听端口 编辑:程序博客网 时间:2024/06/05 20:10

如题,本周课程主要讲述云数据中心的网络管理,和网络基础设施的共享。

Targets and Motivations

具体的目标与动机可以参考下面这篇文章:
VL2: A Scalable and Flexible Data Center Network

我们希望数据中心做到高可伸缩性和高利用率,也就是灵活高效

To be agile and cost eff ective, data centers should allow dynamic resource allocation across large server pools. In particular, the data center network should enable any server to be assigned to any service.

动机:

The motivations for building such shared data centers are both economic and technical: to leverage the economies of scale available to bulk deployments and to benefit from the ability to dynamically reallocate servers among services as workload changes or equipment fails.

要做到数据中心高利用率,关键是灵活,为任意服务分配任意的server:

Agility — the capacity to assign any server to any service

为了实现灵活性,我们希望达到这样的目标:

  • 流量平均
    server到server的网络流量应该只与网卡性能有关,而不应该受到网络拓扑的制约。
  • 独立性能
    server应该表现出独立的性能,也就是说,不会受到其他servers突发吞吐量的影响。例如两个servers在同一个机架上,其中一台server流量非常大,但是另一台server的网络流量却不会因此而受到影响。
  • Layer-2 semantics
    被分配的servers 看起来就像是连在同一个LAN上一样。

VL2 Design

VLB

寻址与路由

为了实现灵活性,使得服务(service)可以在物理数据中心上灵活移动。VL2使用了两类IP地址:1.Locator Addresses; 2.Application Addresses。他们之间可以动态转换。如上图。

Two separate IP address spaces are used – one to identify an application, one to identify its location – with a dynamic translation step between these.

  • Locator Addresses(LAs): 用于识别 rack or top-of-rack (ToR) switch
  • Application Addresses (AAs): 用于识别 server or virtual server

可见,LA是一种物理标识,而AA是一种虚拟标识。当一项服务在VL2内部迁移的时候,它的AA保持不变,但是LA可能会改变。

选路方面,VL2采用类似于ECMP的策略,也就是一种多路径随机选路的方法。这是基于对数据中心的测量结果而选择的策略。测量结果显示:

  • 大量的small flows
  • 并发而且无规律

选择随机选路是因为:

Individual flows may get unlucky by being (randomly) assigned to a path that causes congestion, but if this happens it will be resolved quickly because the flow will last for only a relatively short time.
Other solutions that dynamically change the routing of flows based on current congestion could theoretically perform better. But, it would be difficult to make these schemes practical because the overall network traffic pattern changes very quickly.

0 0