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（二）

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，今天是第二部分。

II. GENERAL MODEL TO ALLOCATE RULES IN OPENFLOW
In this section, we formalize a general optimization model for OpenFlow rule allocation and endpoint policy enforcement.The goal of the optimization is to find an allocation of forwarding rules in an OpenFlow network such that the high-level objectives of the operator are respected and network constraints are satisfied. However, depending on the high-level objectives and the network constraints, it may not be possible to satisfy the endpoint policy for every flow and packets of flows that cannot respect the endpoint policy are then forwarded on an arbitrary default path. In the context of OpenFlow, we assume the existence of: (1) a centralized controller that can be reached from every switch in the network and (2) a default path used in every switch to forward packets that do not match any forwarding rule to the controller.

翻译：

第二部分：在OpenFlow中的规则分配通用模型

在这一部分内容中，我们基于OpenFlow的规则分配与端点规则实施形式化出一个通用的优化模型。这一优化的目标是找到一个在OpenFlow网络中转发规则的分配方式，使得它能符合操作者（规定的）高等级目标和网络自身的限制。然而，基于高等级目标和网络限制来考虑，可能并不是每一条流都能满足端点规则，那些不能遵守端点规则的流量包就（只能）在任意一个默认路径上进行转发。在OpenFlow情境中，我们假设它们的存在：（1）一个能够被网络中所有交换机到达的中心控制器和（2）在每一个交换机中（都存在的），转发那些不能匹配到控制器的任何转发规则的包的一条默认路径。

Based on these assumptions, our optimization model is expressed as an Integer Linear Program (ILP) with constraints and the goal is to maximize an objective function that abstracts the high-level objectives of the operator. Without loss of generality, we assume that one forwarding rule is used for at most one flow. This assumption is also used in [9] to keep the core simple with exact matching rules and easy to manage flows (e.g., rate limitation, accounting). Moreover this has the advantage of keeping our model linear (see V-B). 

翻译：

基于这些假设，我们的最优化模型可以表示为一个带有限制条件并以最大化抽象操作者高等级目标的目标函数为目的的整数线性规划。不失一般性，我们假设一个转发规则最多只能被一条流所使用。这一假设在[9]中也用到了，以保证内核即精确的匹配规则和流管理（速率限制，审计）简单化的简单性。进一步来说，这个假设可以保证我们的模型是线性的（见V-B）。

In the following, we define a flow f ∈ F as a set of packets matching a pattern, starting from one ingress link lf ∈ I and targeting one of the egress links el ∈ E(f). We mean by F the network workload, I the set of ingress links of the network, E(f) included in E is the set of all possible egress links and pf is the packet rate of flow f.

翻译：

接下来，我们定义了一个流f∈F来表示映射到一个（特定）模式的包的集合，它从一条入口链路 lf ∈ I 出发，朝向一条出口链路el ∈ E(f)。其中F代表网络工作负载，I代表网络中入口链路的集合，E(f)包含于E是所有可能出口链路的集合，pf是流f的包速率。

The optimization builds an |F|-by-|L| Boolean allocation matrix denoted by A = (af,l), where af,l indicates whether flow f passes through the directional link l = (u, v); u, v ∈ S+ from node u to node v or not. We refer to Table I for the definition of the different notations used along this paper.

翻译：

这一最优化建立了一个以 A = (af,l)来表示的|F|-|L|的布尔分配矩阵，其中 af,l表示流f是否经过有向链路l= (u, v); u, v ∈ S+（所有结点的集合）。我们用表I中的定义来表示本文章中要用到的一些概念。

Our optimization model is twofold. One part implements the high-level objectives and the other defines the constraints imposed by the network. For the first part, and without loss of generality, the optimization of the high-level objectives can be written as the maximization of an objective function F(A, . . . ). Additional constraints can be added to account for the real network conditions and to limit the space of possible solutions.

翻译：

我们的最优化模型分为两部分。一部分实现高等级目标而另一部分用来定义网络施加的限制。对于第一部分，不失一般性，高等级目标的优化可以写成是一个目标函数F(A, . . . )的最大化。附加限制可以根据实际的网络条件来增添，以限制可能的解空间。

The second part of the model consists of a set of constraints on the allocation matrix A to ensure that network limitations and the endpoint policy are respected. Constraints related to the network are defined so to avoid forwarding loops, bandwidth overload, or memory overflow while endpoint policy constraints ensure that packets can only be delivered to valid egress links.

翻译：

模型的第二部分包含一个分配矩阵A上限制的集合来确保能够符合网络限制和端点规则。网络相关的限制被定义，这样就可以避免转发环路、带宽过载或者是内存溢出，而端点规则限制可以保证包只能被送到有效的出口链路上。

网络限制：

Constraint (1) verifies that af,l is a binary variable. To avoid forwarding loops, acceptable solutions must satisfy flow conservation constraints (2) that ensure that the traffic entering a switch always leaves the switch. Constraint (3) is a sanity constraint. It indicates that among all ingress links, packets of the flow can only traverse the ingress link of f.

翻译：限制（1）说明af,l是一个二进制变量。为了避免转发环路，可接受的解决方案必须满足流保护限制（2），以此来确保一个进入了某个交换机的流量总会离开这个交换机。限制（3）是一个健全限制。它表明在所有的入口链路中，（对于任意流f而言）流包只能穿越f的入口链路。

带宽限制：

Constraint (4) accounts for bandwidth limitation and ensures that the sum of the rates of the flows crossing a link l does not exceed its capacity.

解释：

限制（4）解释带宽限制，其保证了流在一条链路上所通过的速率之和不会超过它的容量。

内存限制：

Constraint (5) accounts for when the memory of each switch is known in advance. On the contrary, when the memory to be allocated on a switch is flexible (e.g., in a Network-as-a-Service context or in virtual private networks where the memory is divided between multiple tenants), the operator may see the memory as a total budget that can be freely divided between switches which is accounted by constraint (6).

翻译：

限制（5）解释当每个交换机的存储能力已经预先知道时的情形。从另一角度来看，当分配在交换机上的内存是灵活的（如在一个服务型网络环境或者是在虚拟私人网络中，内存是被划分给多个租户的），操作者也许会根据限制（6）将内存视为一个可以被自由切分给多个交换机的总的预算。

To route a flow f via a directed link l = (s, d), a rule must be installed on switch s. However, if the next hop dictated by the forwarding rule is the same as the one of the default action of the switch, it is unnecessary to install the rule. This simple aggregation of forwarding rules is taken into account in constraints (5) and (6). We refer to Sec. V-B for a discussion about rule aggregation.

翻译：

为了将一条流f通过一条有向链路 l = (s, d)进行路由，交换机上必须安装一个规则。但是，如果被转发规则所限定的下一跳与交换机的默认行为（之一）相同，安装规则就不是必要的。这一简单的转发规则聚合需要在限制（5）和（6）的情形下加以考虑。我们参考V-B部分来进行规则聚合的讨论。

端点规则限制：

Flows need to satisfy the endpoint policy, i.e., packets of flow f should exit the network at one of the egress points predefined in E(f). However, it may not be possible to allocate each single flow and thus, some will be diverted to the controller instead of their preferred egress point. Constraint (7) and (8) ensure that the endpoint policy is respected by imposing that packets of a flow either exit at one valid egress link or at the controller.

翻译：

流需要满足端点规则，即：流f的包需要通过E(F)中预先定义的出口点之一离开网络。但是，也许不是每一个条简单流都能得到分配，因此有些流就会被转移到控制器而不是原先所偏好的出口点。限制（7）和（8）通过实施流包或者从一个有效出口链路或者在控制器端离开这一策略保证了端点规则能够被满足。

The allocation matrix is a source of information for an operator as it provides at the same time the forwarding table,switch memory occupation, and link usage for a given high-level objective and endpoint policy. It is also important to notice that while a problem may have several equivalent solutions, it may also be unsolvable, depending on the objective function and the constraints. In addition, the general problem is NP-hard as Sec. III-B demonstrates.

翻译：

分配矩阵对于一个操作者而言，是一种信息来源，因为它能同时提供转发表，交换机内存占用情况，以及在给定高等级目标和端点规则下的链路使用情况。还有一点很重要值得注意的是，当一个问题可能有几种等价解决方案时，根据目标函数和限制，它也可能是无解的。另外，这个一般性问题是如第三部分-B里所论证的那样，是NP难求解的。