NaaS: Network-as-a-Service in the Cloud

云计算已经极大地改变了IT领域。今天，小公司甚至个人都可以访问大型数据中心（dc）中几乎无限的资源，以运行计算要求很高的任务。这引发了基于大量数据的“大数据”应用程序的兴起。这些应用包括传统的面向批处理的应用，如数据挖掘、数据索引、日志收集和分析，以及科学应用，以及实时流处理、web搜索和广告。为了支持大数据应用，并行处理系统（如MapReduce）采用了分区/聚合模型：一个大的输入数据集分布在许多服务器上，每个服务器处理一个共享的数据集。然后必须聚合本地生成的中间结果以获得最终结果。分区/聚合模型的一个公开挑战是，当服务器之间交换大量数据流量时，它会导致数据中心中网络资源的高度争用。Facebook报告称，对于26%的处理任务，网络传输占用了50%以上的执行时间。这与其他研究结果一致，表明网络往往是大数据应用的瓶颈。提高数据中心中这种网络绑定应用程序的性能已经引起了研究界的极大兴趣。一类解决方案侧重于通过使用覆盖网络来分发数据和执行部分聚合来减少带宽使用。然而，这需要应用程序对物理网络拓扑进行逆向工程，以优化覆盖网络的布局。即使对物理拓扑有完美的了解，仍然存在根本的低效率：例如，如果服务器只有一个网络接口，那么任何服务器扇形输出高于一个的逻辑拓扑都不能最佳地映射到物理网络。其他建议通过更复杂的拓扑结构或更高容量的网络来增加网络带宽。然而，根据一些估计，新的拓扑和网络过度配置会增加数据中心的运营和资本支出，最多可达5倍，这直接影响到租户成本。例如，Amazon AWS最近引入了具有全等分10gbps带宽的Cluster Compute实例，每小时的成本是默认值的16倍。相反，我们认为，为数据中心租户提供高效、简便和安全的网络操作控制，可以更有效地解决这个问题。我们没有过度供应，而是专注于利用特定于应用程序的知识来优化网络流量。我们将这种方法称为“网络即服务”（NaaS），因为它允许租户自定义从网络接收的服务。支持naas的租户可以部署自定义路由协议，包括多播服务或任意播/即时播协议，以及更复杂的机制，如基于内容的路由和以内容为中心的网络。通过修改路径上数据包的内容，它们可以有效地实现高级的、特定于应用程序的网络服务，如网络内数据聚合和智能缓存。像MapReduce这样的并行处理系统将大大受益，因为数据可以在路径上聚合，从而减少了执行时间。键值存储（例如memcached）可以通过在网络中缓存流行的键来提高性能，与仅终端主机部署相比，这可以减少延迟和带宽使用。NaaS模式通过提高租户应用程序的性能（通过高效的网络内处理），同时降低开发复杂性，有可能彻底改变当前的云计算产品。它旨在将分布式计算和网络通信结合在一个单一的、连贯的抽象中，向“数据中心即计算机”的愿景迈出了重要的一步。

Cloud computing has significantly changed the IT landscape. Today it is possible for small companies or even single individuals to access virtually unlimited resources in large data centres (DCs) for running computationally demanding tasks. This has triggered the rise of "big data" applications, which operate on large amounts of data. These include traditional batch-oriented applications, such as data mining, data indexing, log collection and analysis, and scientific applications, as well as real-time stream processing, web search and advertising.To support big data applications, parallel processing systems, such as MapReduce, adopt a partition/aggregate model: a large input data set is distributed over many servers, and each server processes a share of the data. Locally generated intermediate results must then be aggregated to obtain the final result.An open challenge of the partition/aggregate model is that it results in high contention for network resources in DCs when a large amount of data traffic is exchanged between servers. Facebook reports that, for 26% of processing tasks, network transfers are responsible for more than 50% of the execution time. This is consistent with other studies, showing that the network is often the bottleneck in big data applications.Improving the performance of such network-bound applications in DCs has attracted much interest from the research community. A class of solutions focuses on reducing bandwidth usage by employing overlay networks to distribute data and to perform partial aggregation. However, this requires applications to reverse-engineer the physical network topology to optimise the layout of overlay networks. Even with perfect knowledge of the physical topology, there are still fundamental inefficiencies: e.g. any logical topology with a server fan-out higher than one cannot be mapped optimally to the physical network if servers have only a single network interface. Other proposals increase network bandwidth through more complex topologies or higher-capacity networks. New topologies and network over-provisioning, however, increase the DC operational and capital expenditures-up to 5 times according to some estimates-which directly impacts tenant costs. For example, Amazon AWS recently introduced Cluster Compute instances with full-bisection 10 Gbps bandwidth, with an hourly cost of 16 times the default. In contrast, we argue that the problem can be solved more effectively by providing DC tenants with efficient, easy and safe control of network operations. Instead of over-provisioning, we focus on optimising network traffic by exploiting application-specific knowledge. We term this approach "network-as-a-service" (NaaS) because it allows tenants to customise the service that they receive from the network. NaaS-enabled tenants can deploy custom routing protocols, including multicast services or anycast/incast protocols, as well as more sophisticated mechanisms, such as content-based routing and content-centric networking.By modifying the content of packets on-path, they can efficiently implement advanced, application-specific network services, such as in-network data aggregation and smart caching. Parallel processing systems such as MapReduce would greatly benefit because data can be aggregated on-path, thus reducing execution times. Key-value stores (e.g. memcached) can improve their performance by caching popular keys within the network, which decreases latency and bandwidth usage compared to end-host-only deployments.The NaaS model has the potential to revolutionise current cloud computing offerings by increasing the performance of tenants' applications -through efficient in-network processing- while reducing development complexity. It aims to combine distributed computation and network communication in a single, coherent abstraction, providing a significant step towards the vision of "the DC is the computer".