一个为无阻流量规则分配与端点规则实施而设计的通用最优化框架（一）

写在前面：本文是我做毕业设计参考的一篇并不如何有名的外文文献，其题名为：《OFFICER：A general optimization framework for OpenFlow  Rule Allocation and Endpoint Policy Enforcement》。有兴趣的朋友可以谷歌该题目，找到就可以下载了。不过为了方便，我尽量还是中英文对照着来写吧，今天是第一部分。

OFFICER: A general Optimization Framework for OpenFlow
Rule Allocation and Endpoint Policy Enforcement
Xuan-Nam Nguyen, Damien Saucez, Chadi Barakat, Thierry Turletti
INRIA Sophia Antipolis M´editerran´ee, France
FirstName.LastName@inria.fr

上文翻译略。

Abstract—The Software-Defined Networking approach permits to realize new policies. In OpenFlow in particular, a controller decides on behalf of the switches which forwarding rules must be installed and where. However with this flexibility comes the challenge of the computation of a rule allocation matrix meeting both high-level policies and the network constraints such as memory or link capacity limitations. Nevertheless, in many situations (e.g., data-center networks), the exact path followed by packets does not severely impact performances as long as packets are delivered according to the endpoint policy. It is thus possible to deviate part of the traffic to alternative paths so to better use network resources without violating the endpoint policy. In this paper, we propose a linear optimization model of the rule allocation problem in resource constrained OpenFlow networks with relaxing routing policy. We show that the general problem is NP-hard and propose a polynomial time heuristic, called OFFICER, which aims to maximize the amount of carried traffic in under-provisioned networks. Our numerical evaluation on four different topologies shows that exploiting various paths allows to increase the amount of traffic supported by the network without significantly increasing the path length.

翻译：

摘要——软件定义网络方式允许实现新的规则。尤其是在无阻流量（OpenFlow）中，一个控制器代表那些转发规则必定被安装了的交换机（switch）做出决定。然而伴随着这种可喜的灵活性，也有一个巨大的挑战：对于一个既符合高等级规则又有网络限制（如存储或者链路容量限制）的规则分配矩阵计算（非常复杂）。尽管如此，在很多情形中（例如数据中心网络），只要数据包根据端点规则进行转发，那么包所经由的确切路径并不会严重影响表现性能。因此将一部分流量转移到其他路径以便更好地利用网络资源而同时却不违背端点规则成为了可能。在这一篇文章中，我们提出了一个基于带有松散路由规则的资源受限OpenFlow网络的规则分配问题的线性优化模型。我们证明了这个广义问题是NP难求解的并且提出了一个名为OFFICER的多项式时间启发法，它的目标是最大化（资源）不充分网络中的传输流量的大小。我们对于四种不同拓扑结构的数值估计展示了：利用不同的路径可以增加网络支持的流量大小而不会显著地增加路径长度。

I. INTRODUCTION
The role of a network is to route each packet from an ingress link (i.e., the link from which the packet entered the network) to an egress link (i.e., the link at which the packet leaves the network).1 According to operational and economical requirements, the choice of the egress link to which a packet must be forwarded is dictated by the Endpoint Policy and the actual path followed by a packet in the network is decided by the Routing Policy [1].

翻译：

第一部分：介绍

网络所扮演的角色是将每一个包从一个入口链路（即是包进入到网络的那条链路）路由到一个出口链路（即是包离开网络的那条链路）。1 根据操作和经济需求，一个包必须被转发到的出口链路的选择是由端点规则所指定的，而一个包实际通过的路径是由路由规则所决定的。[1]

Endpoint policies are driven by high-level economical and technical considerations. For example, shared-cost links are often privileged by ISPs and data-centers make sure that packets are delivered to servers able to handle them. On the other hand,routing policies are related to the good use of resources in the network. Shortest-path routing is the most common routing policy. Its advantages stem from the fact that it minimizes the amount of links and nodes traversed by a packet across the network and that routing tables are computed in polynomial time [2] but other routing policies are also possible, for instance, compact routing [3].

翻译：

端点规则是被高等级经济和技术考虑所驱动的。例如，共享消费链路经常被网络服务提供商（ISPs）特权占用，数据中心要保证包能够被传送到可以处理它们的服务商那里去。从另一方面来说，路由规则与网络中资源的良好使用息息相关。最短路径路由是路由规则中最常见的。它的优点在于它最小化了包穿越网络所需的链路和节点的数量，并且路由表是以多项式时间来计算的[2]。但其他路由规则也是可以的，比如密集路由[3]。

From that point of view, respecting the endpoint policy is essential while the routing policy is just a tool to achieve this goal [4]. Unfortunately, relaxing routing policies and removing strong path requirements is not practically doable when the network relies on distributed routing algorithms as it would imply a high signaling overhead to ensure consistency of decisions [5]. But with the advent of Software-Defined Networking (SDN) and OpenFlow in particular, it is now possible to manage routing using a centralized approach without losing in terms of scalability or robustness [6]. OpenFlow allows operators to conceive their network as a black box aiming at carrying packets from sources to destinations [6], [4], [7]. The network thus becomes a single entity that the operator can program instead of a bunch of devices to configure.This is achieved in OpenFlow thanks to a logically centralized that fetches information from the network, computes appropriate routes according to the operator wills and network conditions, and then transparently pushes the corresponding forwarding rules into the switches.

翻译：

从这一观点来看，符合端点规则是非常重要的，而路由规则只是实现这一目标的一个工具[4]。不幸的是，当网络依赖于分布式路由算法时，松散路由规则以及移除强路径需求实际上并不可行，因为那意味着一个高信令开销来保证决定的一致性[5]。但是随着软件定义网络的出现尤其是无阻流量（OpenFlow），现在使用一个中心方式管理路由而不失规模性和强健性是可行的[6]。无阻流量（OpenFlow）允许操作者将网络视为一个旨在将包从从源运输到目的的黑盒子[6],[4],[7]。因此整个网络就成为了一个简单的实体，操作者就可以通过编程（控制）而不是用一堆设备来配置。这在无阻流量（OpenFlow）中之所以行得通，多亏了一个逻辑中心化（控制端），它从网络中取得信息，根据操作者的意愿和网络条件计算合适的路由（通路），然后透明地将相应转发规则放入交换机中。

We illustrate the gain from relaxing routing policy in Fig. 1 that shows a symmetric network of 8 switches with two ingress links (East and West) and two egress links (North and South).In this example, the endpoint policy stipulates that destinations A and B must be reached by the North egress link while any other destination must be reached by the South egress link. With the shortest path routing policy (Fig. 1(a)), every destination is reached in 3 hops and for a total of 15 routing entries. With a policy minimizing the number of routing entries (Fig. 1(b)), the routing table is reduced to 9 entries but the memory reduction comes at the cost of longer paths for A and
B (i.e., 4 hops). However, in practice networks might have bandwidth or memory constraints to be respected. For instance, suppose in our network example that each switch can store 2 routing entries. In this case, the two previous routing policies cannot be applied as they would violate the constraints whereas Fig. 1(c) shows an allocation that respects both the endpoint policy and the switches’ constraints.

翻译：

我们通过如图1所示的由8个交换机，2条入口链路（东西向）和2条出口链路（南北向）组成的对称网络来阐述松散路由规则的收益。在这个例子中，端点规则要求目的A和B必须通过北出口链路而其他目的地必须通过南出口链路到达。使用最短路径路由规则（如图a）时，每个目的都可以通过3跳到达，总共有15个路由条目。使用最小化路由表项规则（如图b）时，路由表就被缩减到只剩下9个条目但是存储的减少代价是到A、B的路径更长（4跳）。然而，在实际网络中可能有带宽和存储限制，这需要引起重视。例如，假设在我们的网络中每个交换机可以存储2个路由条目。在这一情形下，之前的两个路由规则就不能再适用了，因为它们会违背限制条件。而图c给出了一个分配方案，它能够同时满足端点规则和交换机限制。

Departing from the flexibility offered by OpenFlow, we present OFFICER, a general algorithm to calculate and implement efficient forwarding rules in switches. OFFICER treats the network as a black box that must satisfy the endpoint policy imposed by the operator and tries to get the maximum from the available resources by adapting the routes followed by the different packets towards their desired egress links. When the network is under-provisioned, least valuable packets are routed through a default slow path designed to minimize resource usages. As suggested in [4] and [7], we believe that in most networks, enforcing a particular path is not necessary as long as the endpoint policy is respected. Actually, not relying on strict routing policies allows better utilization of the network capacity, reducing so bandwidth wastage and congestion events [4].Relaxing routing policy is particularly useful in case of scarce network resources as shown in Fig. 1 and in [7].

翻译：

从无阻流量（OpenFlow）所提供的灵活性出发，我们提出了OFFICER这个通用算法来计算和实现在交换机中有效的转发规则。OFFICER 将网络视为一个必须满足由操作者限定的端点规则的黑盒子，并且尝试通过适应不同包朝理想的出口链路时经过的路线来从可达资源中取得最大值（即效益最大化）。当网络是（资源）不充分时，最少的高价值包是通过为最小化资源使用（率）而设计的默认慢路径路由的。如同[4]和[7]中所提的那样，我们相信在绝大多数网络中，只要符合端点规则，那么强制使用一个特殊的路径就是不必要的。事实上，不依赖于严格的路由规则可以使网络容量得到更好的利用，因此减少了带宽浪费和冲突事件[4]。松散路由规则在在如图1与[7]中所示的资源缺乏网络中尤其有用。

The remaining of this paper presents our algorithm OFFICER to allocate forwarding rules in OpenFlow networks. This algorithm is the result of a general ILP optimization model formulated in Sec. II, where the OpenFlow network is modeled as a directed graph interconnecting switches and the rules to install on switches are to be found. Our model is general in the sense it can accept any endpoint policies and can accommodate any reward functions (i.e., high-level objective) that the operator aims to maximize. Its novelty can be summarized in two main points: (i) modeling the network as a black box respecting he endpoint policy, and (ii) getting the maximum from the available resources by relaxing routing policy, the rest of the traffic that cannot be installed is routed on a default path.As to be discussed in the related work section (Sec. VI), we are the first to propose a solution making such abstraction of an OpenFlow network, with a clear gain in terms of the volume of traffic that can be correctly assigned to its desired egress point. To illustrate the flexibility of our proposition, we study the particular case of network that is missing memory to
account for all forwarding rules in Sec. III. This is a growing problem in networks because of the increase in size of routing tables but also due to the trend to keep outdated routers in operation [8]. This problem can even be exacerbated with OpenFlow as it enables very fine granularity on forwarding decisions. In Sec. IV we numerically evaluate the costs and benefits of relaxing routing policy on ISP and data-center topologies and present different heuristics that approximate the optimal algorithm in polynomial time. We open some discussion in Sec. V and consider the related work in Sec. VI to finally conclude in Sec. VII.

翻译：

本文的剩余部分将会展示我们的OFFICER算法，看它如何在OpenFlow网络中分配转发规则。这一算法是一个在第二部分中规范化的通用ILP最优化模型（启发）下产生的，在那一部分中，OpenFlow网络被模型化为一个交换机相互连接的有向图，而安装在交换机上的规则亟待发现。我们的模型是通用的，因此它能够容纳任何的端点规则，也能适应任何操作者想最大化的的收益函数（高等级目标）。它的新颖性可以归纳为以下两点：第一，将网络模型化为一个符合端点规则的黑盒子；第二，通过松散路由规则从可达资源中获取最大值，而其余（转发规则）不能安装的流量就通过默认路径路由。后面将在相关工作部分（第六部分）中讨论到，我们是第一个提出抽象化OpenFlow网络的解决方案的，这一方案对于能够被正确指派到所需的出口点的流量体积做了清晰的收益说明。为了阐述我们提案的灵活性，我们研究一种特别的网络案例，即失去存储（空间）导致了第三部分中所有转发规则的这一例子。这在网络中是一个日渐严峻的问题，不仅是因为路由表体量变大，还因为仍然将过时路由器放置在网络中进行操作的趋势。这个问题在OpenFlow环境下会更严重，因为它能够保持转发决策的良好颗粒性。在第四部分，我们对ISP和数据中心拓扑上的基于松散路由规则的开销和收益进行了数值估计，并且提出了逼近多项时间下的理想算法的不同启发式算法。我们将在第五部分作出一些开放式的讨论，在第六部分考虑相关的工作并于第七部分做出最后的总结。