一个为OpenFlow规则分配与端点规则实施而设计的通用最优化框架（五）

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，今天是第五部分。

V. DISCUSSION
With this section we provide a broad discussion on the model presented in Sec. II  as well as the assumptions that drove it.
A. Routing policy
Relaxing routing policy allows better usage of the network but comes with the expense of potential high path stretch.Nevertheless, nothing prevents to add constraints in our model to account for a particular routing policy. For example, the constraint can be added to control the maximum path length of each flow. This constraint binds the path length to an arbitrary value pre-computed by the operator, withα(f) : F->R. For example,α(f) = h · shortest_path_length(f)  to authorize a maximum path stretch h (e.g., h = 1.5 authorizes paths to be up to 50% longer than the corresponding shortest paths).
翻译：

V.讨论

在这一部分我们对第二部分所展示的模型和驱动它的假设提供一个大概的讨论。

A.路由规则

松散路由规则允许对网络更好的使用但也伴随着潜在的高路径延展这一代价。尽管如此，没有什么能够阻止向我们的模型中加入限制来对一个特殊的路由规则作出解释。例如，可以加入限制来控制每条流的最大路径长度。这一限制绑定了路径长度和操作者预先计算的任意值，通过式子α(f) : F->R。例如，

α(f) = h · shortest_path_length(f) 允许一个最大路径延展h（例如，h = 1.5就允许路径可以比相应的最短路径长百分之五十）。

B. Rule Aggregation
To aggregate two rules having the same forwarding action into one single rule, a common matching pattern must be found between the two rules. Constraints (5) and (6) provide a first step towards rules aggregation: on a switch, if the forwarding decision for a flow is the same as the default action, the rule for the flow does not need to be installed. However, a problem occurs when the common matching pattern also matches for another rule that has a different action. The latter rule should not be covered by the aggregating rule as that could create loop events or incorrect forwarding. Consequently, the construction of the minimal set of rules in a switch by using aggregation requires the knowledge of the allocation matrix that, in turn, will be affected by the aggregation. This risk of non-linearity is a reason why we assume that one forwarding rule is used for at most one flow and why we limit aggregation to the default rule only.
翻译：

B.规则聚合

为了将两个具有同样转发动作的规则聚合成一个规则，就必须在这两个规则之间建立一个通用的匹配模式。限制（5）和（6）提供了规则聚合的第一步：在一个交换机上，如果一个流的转发决定与默认动作是相同的，对于这条流的这一规则就可以不必安装了。然而，当这个通用模式也匹配到另一个具有不同动作的规则时，问题就出现了。后面的规则就不应该被聚合的规则所覆盖，因为那样会导致循环事件或者不正确的转发。因此，使用聚合建立的最小规则集合需要这一知识，分配矩阵会反过来被聚合所影响。非线性的风险就是为什么我们假设一个转发规则最多只被一条流使用以及为什么我们要限制仅仅聚合到默认规则的一个理由。

C. Multipath
The model presented in Sec. II  assigns one forwarding path per flow. As a result, all the packets of a flow follow the same path to the egress link, which ensures that packet arrival order is maintained. Nevertheless, our model does not prevent multipath routing. To do so, the pattern matching of a flow to be forwarded on several paths must be redefined from the one used in case of one forwarding path. From a network point of view, the flow will then be seen as multiple flows, one per matching pattern. Consequently, the optimizer might give different forwarding paths for packets initially belonging to the same flow. For example, one can assign a label to packets when they enter the network and then use labels to decide to which rule the packet matches. This may increase significantly the number of rules to be installed in the network and the gain of having several such paths must be compared to the cost of having them. In most situations, multipath routing at the flow level might not be necessary as we are not enforcing any routing policy in our model, which limits the risk of having the traffic matching one rule to be enough to saturate one link.

翻译：

C.多径

第二部分中提出的模型对每一条流指定一个转发路径。结果，一条流上的所有包都会经由相同的路径到达出口链路，这就保证了包的到达顺序是可以维持不变的。尽管如此，我们的模型不阻止多径路由。为了能够那么做，在几条路径上被转发的一条流的模式匹配必须根据已使用的一条转发路径的案例重新定义。从一个网络的视角去看，这条流将会被视为多条流，每个都有一个（不同的）匹配模式。因此，最优化可以对最初属于同一条流的包给出不同的转发路径。例如，当包进入网络的时候，可以给它们分配一个标签，然后利用标签来决定包匹配到哪一个规则。这一举措会显著增加安装在网络中的规则数量，而新增多条路径的收益必须与付出的开销相比较。在大多数情形中，在流等级进行的多径路由也许不是必要的，因为我们在模型中并没有强制实施任何路由规则，这也就限制了具有流量匹配到一个规则就足以使一条链路饱和的风险。

VI. RELATED WORK
Rule allocation in OpenFlow has been largely covered over the last years. Part of the related work proceeds by local optimization on switches to increase their efficiency in handling the installed rules. The other part, which is more relevant to our work, solves the problem network-wide and produces a set of compressed rules together with their placement. Our present research builds upon this rich research area and presents an original model, together with its solution, for the rule allocation problem where the routing can be relaxed for the only objective of placing as many as rules as possible that respect the predefined endpoint policy.

翻译:

VI.相关工作

OpenFlow中的规则分配在过去几年间被大量讨论到。一个相关的工作是对交换机做本地优化，以增加它们在处理已安装的规则上的效率。另一个与我们工作更相关的是，解决全网络问题以及根据它们的放置位置生成一个压缩的路由集合。我们现在的研究是建立在这一繁荣的研究领域并且提出了一个对唯一目标（可这个目标上放置任意多的规则，只要它们符合预先定义的端点规则）路由可松散的规则分配问题的原始模型和解决方案。

For the first part, several mechanisms based on wildcard rules have been proposed to minimize the rule space consumption on switches as well as to limit the signaling overhead between switches and controller. DevoFlow [18] uses wildcard rules to handle short flows locally on switches. DomainFlow [19]
divides the network into one domain using wildcard rules and another domain using exact matching rules. SwitchReduce [20] proposes to compress all rules that have the same actions into a wildcard rule with the exception of the first hop switch.

翻译：

对于第一部分，几种建立在通配符规则之上的机制已经被提出，用来最小化交换机上的规则空间消耗和交换机与控制器之间的信令开销。DevoFlow[18]使用通配符规则在交换机本地处理短流。DomainFlow[19]将网络用通配符规则划分为一个域，用精确匹配规则分划为另一个域。SwitchReduce[20]提出了压缩所有具有相同动作的规则到一个通配符规则中，只是第一跳交换机例外。

To reduce further memory usage, latest versions of OpenFlow support pipelining and multi-level flow tables [21]. Consequently,the large forwarding table is split in a hierarchy of smaller tables that can be combined to build complex forwarding rules with less entries. However, even though these techniques improve memory usage, they do not remove the exponential growth of state with the number of flows and nodes in the network.
翻译：

为了进一步减小内存使用，最新的OpenFlow技术支持流水线和多级流表[21]。因此，一个大的转发表就可以被切分成多个更小的表，可以合并它们来建立具有更少条目的复杂转发规则。然而，即使这些技术提升了内存使用，但它们它们不能消除随着网络中流和节点数的增加时呈指数增长的状态。

As for the second part, some works suggest to use special devices to perform rule placement. DIFANE [22] places the most important rules at some additional devices, called authority switches. Then, ingress switches redirect unmatching packets towards these specific devices, which enables reducing load on the controller and, at the same time, decreasing the number of rules required to be stored on ingress switches. vCRIB [23] installs rules on both hypervisors and switches to increase performance while limiting resource usage. Other works optimize rule allocation on switches themselves. Palette [10]
and OneBigSwitch [1] produce the aggregated rule sets that satisfy the endpoint policy and place them on switches while respecting the routing policy and minimizing the resources.However both Palette and OneBigSwitch cannot be used in scenarios where resources are missing to satisfy the endpoint
policy. In [24], the rule allocation is modeled as a constrained optimization problem focusing on the minimization of the overall energy consumption of switches. Finally, the authors in [7] propose a network-wide optimization to place as many rules as possible under memory and link capacity constraints.

翻译：

对于第二部分，一些工作提出使用特殊设备来进行规则放置。DIFANE[22]将最重要的规则放置在附加的叫权威交换机的设备上。这样，入口交换机就会将那些没有匹配的包指向这些特殊设备，这就减小了控制器上的负载，与此同时，也减小了需要安装在入口交换机上的规则数量。vCRIB[23]在管理程序和交换机上都安装规则，在减小资源使用的同时增加性能。其他工作则优化交换机本身上的规则分配。Palette[10]和OneBigSwitch[1]生成了能够满足端点规则的聚合规则集合并把它们放置在交换机上，同时也能符合路由规则和最小化资源。然而无论是Palette还是OneBigSwitch都不能在失去资源以满足端点规则的场景中使用。在[24]中，规则分配被模型化为一个聚焦于最小化交换机上总的能源消耗的受限最优化问题。最后，[7]中的著者提出了一个全网络优化来在内存和链路容量限制条件下放置尽可能多的规则。

While the related works presented above focus on particular aspects of the rule allocation problem in OpenFlow, with OFFICER we are the first to propose a general solution that is able to cope with endpoint and routing policies, network constraints, and high-level operational objectives.

翻译：

虽然以上列出的相关工作关注了OpenFlow中规则分配问题的一些特殊角度，但我们的OFFICER却是第一个提出能够处理端点和路由规则，网络限制和高等级操作者目标的一般性解决方案。

VII. CONCLUSION
We presented in this work a new algorithm called OFFICER for rule allocation in OpenFlow. Starting from a set of endpoint policies to satisfy, OFFICER respects as many of these policies as possible within the limit of available network resources both on switches and links. The originality of OFFICER is in
its capacity to relax the routing policy inside the network for the objective of obtaining the maximum in terms of endpoint policies. OFFICER is based on an integer linear optimization model and a set of heuristics to approximate the optimal allocation in polynomial time. The gain from OFFICER was shown by numerical simulations over realistic network topologies and traffic traces.
翻译：

VII.总结

我们在这次工作中提出了一个针对OpenFlow的规则分配问题的新算法OFFICER。从满足一个端点规则的集合开始，OFFICER在无论是交换机还是链路上可能的网络资源受限条件下，能够符合尽量多的这些规则。OFFICER的独创性在于依据端点规则，为获得最大化的目标，在网络中松散路由规则的容量。OFFICER基于整数线性优化和一个在多项时间逼近最优分配的一套启发式算法。OFFICER的收益是通过现实网络拓扑结构和流量追踪来进行数值仿真而展现的。

ACKNOWLEDGMENT
This work is funded by the French ANR under the “ANR-13-INFR-013” DISCO project on SDN.

此处翻译略。

文末附上相关的参考文献。

REFERENCES
[1] N. Kang, Z. Liu, J. Rexford, and D. Walker, “Optimizing the ’One Big Switch’ Abstraction in Software-Defined Networks,” in ACM CoNEXT,Dec. 2013.
[2] R. Bellman, “On a Routing Problem,” Quarterly of Applied Mathematics,vol. 16, pp. 87–90, 1958.
[3] M. Thorup and U. Zwick, “Compact routing schemes,” in SPAA, 2001.
[4] S. e. a. Jain, “B4: Experience with a globally-deployed software defined WAN,” in ACM SIGCOMM, 2013.
[5] J. Perry, A. Ousterhout, H. Balakrishnan, D. Shah, and H. Fugal, “Fastpass: A Centralized Zero-Queue Datacenter Network,” in ACM SIGCOMM, August 2014.
[6] N. McKeown, T. Anderson, H. Balakrishnan, G. Parulkar, L. Peterson,J. Rexford, S. Shenker, and J. Turner, “OpenFlow: enabling innovation in campus networks,” SIGCOMM CCR, vol. 38, no. 2, Mar. 2008.
[7] X. N. Nguyen, D. Saucez, C. Barakat, and T. Turletti, “Optimizing Rules Placement in OpenFlow Networks: Trading Routing for Better Efficiency,”in ACM HotSDN, Apr. 2014.
[8] H. Ballani, P. Francis, T. Cao, and J. Wang, “Making routers last longer with viaggre,” in USENIX NSDI, Berkeley, CA, USA, 2009, pp. 453–466.
[9] X. Jin, H. Liu, R. Gandhi, and S. Kandula, “Dynamic Scheduling of Network Updates,” in ACM SIGCOMM, 2014.
[10] Y. Kanizo, D. Hay, and I. Keslassy, “Palette: Distributing tables in software-defined networks,” in INFOCOM, Apr. 2013, pp. 545–549.
[11] T. Benson, A. Akella, and D. a. Maltz, “Network traffic characteristics of data centers in the wild,” IMC, 2010.
[12] N. Spring, R. Mahajan, and D. Wetherall, “Measuring ISP topologies with Rocketfuel,” SIGCOMM CCR, vol. 32, no. 4, pp. 133–145, 2002.
[13] M. Al-Fares, A. Loukissas, and A. Vahdat, “A scalable, commodity data center network architecture,” ACM SIGCOMM, 2008.
[14] “Abilene,” http://http://sndlib.zib.de.
[15] L. Saino, C. Cocora, and G. Pavlou, “A Toolchain for Simplifying Network Simulation Setup,” in SIMUTOOLS, 2013.
[16] P. Wette, A. Schwabe, F. Wallaschek, M. H. Zahraee, and H. Karl, “MaxiNet: Distributed Emulation of Software-Defined Networks,” in IFIP Networking Conference, 2014.
[17] B. Heller, R. Sherwood, and N. McKeown, “The controller placement problem,” in ACM HotSDN, 2012, pp. 7–12.
[18] A. R. Curtis, J. C. Mogul, J. Tourrilhes, P. Yalagandula, P. Sharma, and S. Banerjee, “DevoFlow: scaling flow management for high-performance
networks,” SIGCOMM CCR, vol. 41, no. 4, pp. 254–265, Aug. 2011.
[19] Y. Nakagawa, K. Hyoudou, C. Lee, S. Kobayashi, O. Shiraki, and T. Shimizu, “Domainflow: Practical flow management method using multiple flow tables in commodity switches,” in ACM CoNEXT, 2013.
[20] A. Iyer, V. Mann, and N. Samineni, “Switchreduce: Reducing switch state and controller involvement in openflow networks,” in IFIP Networking
Conference, 2013.
[21] “OpenFlow Switch Specification,” http://www.opennetworking.org/.
[22] M. Yu, J. Rexford, M. J. Freedman, and J. Wang, “Scalable flow-based networking with DIFANE,” SIGCOMM CCR, 2010.
[23] M. Moshref, M. Yu, A. Sharma, and R. Govindan, “Vcrib: Virtualized rule management in the cloud,” in USENIX HotCloud, 2012.
[24] F. Giroire, J. Moulierac, and T. K. Phan, “Optimizing Rule Placement in Software-Defined Networks for Energy-aware Routing,” in IEEE GLOBECOM, 2014.

后话：总算翻译完了，但是因为后面一部分是即时翻译，肯定有很多内容不甚准确，有兴趣阅读的朋友也不要太指望我的翻译功底啦~

不过，这篇文章有些地方用语确实还比较晦涩难懂，还得再深刻阅读一下才成~

加油！加油！加油！