SIGCOMM12论文简要选读

以下从研究的角度对SIGCOMM 2012上发表的论文有重点的加以总结：

Session A Middleboxes and middleware

1，

Ali Ghodsi, Vyas Sekar, Matei Zaharia, Ion Stoica: Multi-resourcefair queueing for packet processing. 1-12
这篇论文主要针对的是不同middleware对硬件上的资源有不同的需求，原先的研究主要考虑如何对单一资源（如bandwidth）进行优化调度，本文认为各种资源之间是有联系的，通过实验说明一般的针对某一种资源进行调度的算法不能很好的对多种资源进行调度，所以本文提出了一种新的综合对多种资源进行统一调度的方法。

摘要：“Middleboxes are ubiquitous in today’s networks and perform a variety of important functions, including IDS, VPN, firewalling, and WAN optimization. These functions differ vastly in their requirements for hardware resources ( e.g.,CPU cycles and memory bandwidth). Thus, depending on the functions they go through, different flows can consume different amounts of a middlebox’s resources. While there is much literature on weighted fair sharing of link bandwidth to isolate flows, it is unclear how to schedule multiple  resources in a middlebox to achieve similar guarantees. In this paper, we analyze several natural packet scheduling algorithms for multiple resources and show that they have undesirable properties. We propose a new algorithm, Dominant Resource Fair Queuing (DRFQ), that retains the attractive properties that fair sharing provides for one resource. In doing so, we generalize the concept of virtual time in classical fair queuing to multi-resource settings. The resulting algorithm is also applicable in other contexts where several resources need to be multiplexed in the time domain.”

2，Justine Sherry, Shaddi Hasan, Colin Scott, Arvind Krishnamurthy, Sylvia Ratnasamy, Vyas Sekar:  Making middleboxes someone else's problem: network processing as a cloud service. 13-24（系统实现）

本文主要的动机是考虑到middlebox处理服务的复杂性以及昂贵，结合当前火热的云计算技术，将这些作为一种服务转移到云平台上去。但实际中主要的困难在于如何保证这一服务转移后的功能可靠性而又不牺牲性能不增加网络的复杂性。本文的工作中通过实际的进行相关调研，提出了并实现了一种具体的解决方案，而实际的系统效果显示其性能是可以的。

摘要：“Modern enterprises almost ubiquitously deploy middlebox processing services to improve security and performance in their networks. Despite this, we find that today’s middlebox infrastructure is expensive, complex to manage, and creates new failure modes for the networks that use them. Given the promise of cloud computing to decrease costs, ease management, and provide elasticity and fault tolerance, we argue that middlebox processing can benefit from outsourcing the cloud. Arriving at a feasible implementation, however, is challenging due to the need to achieve functional equivalence with traditional middlebox deployments without sacrificing performance or increasing network complexity.

In this paper, we motivate, design, and implement APLOMB, a practical service for outsourcing enterprise middlebox processing  to the cloud. Our discussion of APLOMB is data-driven, guided by  a survey of 57 enterprise networks, the first large-scale academic  study of middlebox deployment. We show that APLOMB solves real  problems faced by network administrators, can outsource over 90%  of middlebox hardware in a typical large enterprise network, and,  in a case study of a real enterprise, imposes an average latency  penalty of 1.1ms and median bandwidth inflation of 3.8%。”

3，

• Robert Escriva, Bernard Wong, Emin Gün Sirer:  HyperDex: a distributed, searchable key-value store. 25-36

对这篇文章的内容比较陌生，主要是介绍对当前NoSQL数据库系统本地与实际远程database之间某一种提高实时性能的方法进行进一步的改进的研究。“The recent trend toward NoSQL data stores has been fueled by scalability and performance concerns at the cost of functionality. HyperDex bridges this gap by providing additional functionality without sacrificing scalability or performance.” 

摘要：“Distributed key-value stores are now a standard component of high-performance web services and cloud computing applications. While key-value stores offer significant performance and scalability advantages compared to traditional databases, they achieve these properties through a restricted API that limits object retrieval—an object can only be retrieved by the (primary and only) key under which it was inserted. This paper presents HyperDex, a novel distributed key-value store that provides a unique search primitive that enables queries on secondary attributes. The key insight behind HyperDex is the concept of hyperspace hashing in which objects with multiple attributes are mapped into a multidimensional hyperspace. This mapping leads to efficient implementations not only for retrieval by primary key, but also for partially-specified secondary attribute searches and range queries. A novel chaining protocol enables the system to achieve strong consistency, maintain availability and guarantee fault tolerance. An evaluation of the full system shows that HyperDex is 12-13×faster than Cassandra and MongoDB for finding partially specified objects. Additionally, HyperDex achieves 2-4 × higher throughput for get/put operations“

session 2 Wireless communication

4 

• Steven Siying Hong, Jeffrey Mehlman, Sachin Rajsekhar Katti: Picasso: flexible RF and spectrum slicing. 37-48
  
  
  （偏硬件，偏通信、偏系统）

• 主要是介绍了如何多种设备同时在同一spectrum上可以同时进行收发数据而不互相干扰的系统方法及其实现。

• 摘要：略

5，

• Jonathan Perry, Peter Iannucci, Kermin Fleming, Hari Balakrishnan, Devavrat Shah
  : Spinal codes. 49-60
  （新编码系统实现）
  介绍一种新的rateless coding系统，这种系统的优点是可以在时变channel下仍然正常的传输数据。
  摘要：“pinal codes are a new class of rateless codes that enable wireless networks to cope with time-varying channel conditions in a natural way, without requiring any explicit bit rate selection. The key idea in the code is the sequential application of a pseudo-random hash function to the message bits to produce a sequence of coded symbols for transmission. This encoding ensures that two input messages that differ in even one bit lead to very different coded sequences after the point at which they differ, providing good resilience     to  noise and bit errors. To decode spinal codes, this paper develops an approximate maximum-likelihood decoder, called the bubble decoder, which runs in time polynomial in the message size and achieves the Shannon capacity over both additive white Gaussian noise (AWGN) and binary symmetric channel (BSC) models. Experimental results obtained from a software implementation of a linear-time decoder show that spinal codes achieve higher throughputthan fixed-rate LDPC codes [11], rateless Raptor codes [33], and the layered rateless coding approach [8] of Strider [12], across a range of channelconditions and message sizes. An early hardware prototype that can decode at 10 Mbits/s in FPGA demonstrates that spinal codes are a practical construction.“

6，

• Jue Wang, Haitham Hassanieh, Dina Katabi, Piotr Indyk: Efficient and reliable low-power backscatter networks. 61-72

不太了解，侧重于编码系统的应用）

针对像RFID这种通过反向传播的系统（如RFID tags需要reader提供能源）中冲突以及对低能耗的需求提出了一种新的方法。并建立一种较为巧妙的模型，将可能的冲突看作一种rateless code从而使得一定程度上仍然能区分出发生冲突的信息。

摘要：“There is a long-standing vision of embedding backscatter nodes like RFIDs into everyday objects to build ultra-low power ubiquitous networks. A major problem that has challenged this vision is that backscatter communication is neitherreliable norefficient. Backscatter nodes cannot sense each other, and hence tend to suffer from colliding transmissions. Further, they are ineffective at adapting the bit rate to channel conditions, and thus miss opportunities to increase throughput, or transmit above capacity causing errors.

• This paper introduces a new approach to backscatter communica tion. The key idea is to treat all nodes as if they were a single virtual  sender. One can then view collisions as a code across the bits trans mitted by the nodes. By ensuring only a few nodes collide at any  time, we make collisions act as a  sparse code  and decode them us ing a new customized compressive sensing algorithm. Further, we  can make these collisions act as a  rateless code  to automatically  adapt the bit rate to channel quality –i.e., nodes can keep collid ing until the base station has collected enough collisions to decode.  Results from a network of backscatter nodes communicating with  a USRP backscatter base station demonstrate that the new design  produces a 3.5  ×  throughput gain, and due to its rateless code, re duces message loss rate in challenging scenarios from 50% to zero.”

session 3 Data centers: latency

• 
  
  7，Balajee Vamanan, Jahangir Hasan, T. N. Vijaykumar: Deadline-aware datacenter tcp (D2TCP). 115-126
  
  
• 8，Chi-Yao Hong, Matthew Caesar, Brighten Godfrey: Finishing flows quickly with preemptive scheduling. 127-138
• 9，David Zats, Tathagata Das, Prashanth Mohan, Dhruba Borthakur, Randy H. Katz: DeTail: reducing the flow completion time tail in datacenter networks. 139-150

这几篇文章都是针对互联网服务中的时延问题加以研究的。现今的互联网上许多的Web service等许许多多的服务都是需要内部多个数据中心提供支持的，而人们在享受服务时对网站的时延有很高的需求。这就需要设计算法来减少多flow下的竞争时间等。

session 4 Measuring Networks

10, Gonca Gürsun, Natali Ruchansky, Evimaria Terzi, Mark Crovella: Inferring visibility: who's (not) talking to whom? 151-162

原以为这篇文章会比较有价值，可能可以在匿名网络中通过traffic 的分析来预测得到有哪些终端在进行通信从而影响安全性。但是仔细看这篇论文会发现：1，这篇文章的网络是假定在目前的AS的BGP之间的主干网络，2，它的目标是让一个AS知道有那些route通过它自己的网络，而不是在网络内部进一步分析匿名传输flow，3它的方法是基于一个观察，即常常一群（源，终）pair具有相同的route路径。所以这篇文章的内容也就不那么吸引我们了。

11， Bernhard Ager, Nikolaos Chatzis, Anja Feldmann, Nadi Sarrar, Steve Uhlig, Walter Willinger: Anatomy of a large european IXP. 163-174

本文主要对于当前Internet的主干网上IXP的流量特点进行了调研，通过分析大量的历史数据，观察到一些与人们预期不太一样的结果。

12,

• Vacha Dave, Saikat Guha, Yin Zhang: Measuring and fingerprinting click-spam in ad networks. 175-186

这篇文章主要介绍了在网络广告服务中，广告投放者如何能解决click-spam问题，即不被虚假的广告点击欺骗，从而不损失自己的广告投送费用。这里没有什么privacy问题，同时由于这种click-spam技术是由不良的程序自身造成的，在http层加以欺骗，也不好通过什么密码、网络安全方法去解决。但似乎这里存在着game，不同的content publisher、ad network、advertiser之间都有自己的利益，同时又是相互存在利益上的依赖与约束。与之有一定联系的是SIGCOMM13上的一篇关于广告的文章，则是介绍publisher如何通过auction来分配自己的广告位，使得参与竞争的人和publisher自身利益都不受损。

session 5： Data center：Resource Management

13，Lucian Popa, Gautam Kumar, Mosharaf Chowdhury, Arvind Krishnamurthy, Sylvia Ratnasamy, Ion Stoica:FairCloud: sharing the network in cloud computing. 187-198

这篇文章挺重要，提出了Data center中会存在如何对network bandwidth进行公平合理分配的问题，主要是三个目的：mini-gurantee、payment proporionality、high utilization这三者之间存在这不可兼得的tradeoff。作者也提出了一些方案解决这一问题。之后我们看到在INFOCOM13上有其他人跟进采用game加以研究的论文。在ICNP13上也有相关论文。值得注意！

14，Di Xie, Ning Ding, Y. Charlie Hu, Ramana Rao Kompella: The only constant is change: incorporating time-varying network reservations in data centers. 199-210

可见本文也研究的是Data center中的network资源分配问题。只不过是从如何更加合理的对这些资源进行abstraction从而更好的方便tenant进行使用这一角度。

15，Peter Xiang Gao, Andrew R. Curtis, Bernard Wong, Srinivasan Keshav: It's not easy being green. 211-222

第一次具体看到有关green computing的论文。这篇论文针对的是分布式data centers中的最优用户request的路由问题。这里作者不仅考虑了时延，还考虑了电力消耗、碳排放等其他方面的开销。这一问题的由来自然是以为数据中心常常为了保证低时延而在物理上分布的十分广阔，而数据中心的用电量又是非常的惊人。

session 6： wireless and mobile networking

16，

• Asaf Cidon, Kanthi Nagaraj, Sachin Katti, Pramod Viswanath: Flashback: decoupled lightweight wireless control. 223-234
  
  又是一篇偏硬件实现的文章，本文作者针对wifi网络中wifi缺乏一个control plane导致系统中许多功能（如公平的资源分配等）难以实现的问题提出了一种新的技术。他们使用FPGA具体实现了他们的技术。

17，Hariharan Shankar Rahul, Swarun Kumar, Dina Katabi: JMB: scaling wireless capacity with user demands.235-246

本文考虑的是wifi这种AP网络中在device用户比较多的情况下，AP的throughput无法满足用户们的要求，认为目前这一困境的主要原因在于网络中用户数据量的scale是与终端device的数目成正比的，而无线网络的throughtput却无法简单的靠增加AP数目得到提升。所以作者在这篇文章中提出了一种新的使用beamforming技术的方法来尝试克服这一困难。

18，Sangtae Ha, Soumya Sen, Carlee Joe-Wong, Youngbin Im, Mung Chiang: TUBE: time-dependent pricing for mobile data. 247-258

本文针对目前ISP对流量的收费仍不合理以及不能较好的处理一天中peak time的网络网络服务质量问题。将Time-dependent pricing这种理念通过技术手段加以实现，分别在ISP端和用户device端提供API，可以让ISP根据其网络的拥堵情况指定不同时段的价位，在用户端则让用户选择在那一个时段进行一些服务。

19，Swarun Kumar, Lixin Shi, Nabeel Ahmed, Stephanie Gil, Dina Katabi, Daniela Rus: CarSpeak: a content-centric network for autonomous driving. 259-270

又是一篇Katabi参与的文章。。。这篇文章主要考虑如何在自动驾驶的汽车之间建立通讯网络，这样的话，各辆车之间就可以通过这一网络交互各自感知到的路障等信息，从而使得自动驾驶汽车可以更加安全，更加快速的对障碍物作出反应。他们还通过具体的无人驾驶汽车的实验以及机器人参与的模拟实验对效果进行了测试。。。。土豪啊。。。

session 7 Network formalism and algorithmics

20，

• Florin Ciucu, Jens Schmitt: Perspectives on network calculus: no free lunch, but still good value. 311-322
  
  

对这个研究内容十分陌生，似乎network calculus是作为排队论在Internet这种环境下的一种新分支，因为Internet中的统计到达模型不再是泊松分布形式的，所以研究得到了network calculus这种理论。本文的内容否定了之前的一个研究结论，并且通过具体的实验支持了他的新论点。

21，

• Mark Reitblatt, Nate Foster, Jennifer Rexford, Cole Schlesinger, David Walker: Abstractions for network update. 323-334
  
  

在SDN等网络中，需要对网络的配置进行更新。传统上，在更新的过程中，即使可以保证更新最初或者更新完成后的最终状态是正确的，但是对于具体更新进行中这一时段内网络的状态则没有保证。本文提出了一种新的方法，使得在更新的整个过程中，网络的配置状态都是可控的，都是有保障的。

22，Yadi Ma, Suman Banerjee: A smart pre-classifier to reduce power consumption of TCAMs for multi-dimensional packet classification. 335-346

本文考虑TCAM在进行路由分包时，常常工作量巨大，十分的消耗能量，服务质量不够好。本文的方法是在进行分类之前进行一个小步骤，对于进入的大量package进行简单的处理，使得后续的路由分包速度得到极大的提升。

 Session 8: Streaming and Content Networking

• • 23
    
  Chen Tian, Richard Alimi, Yang Richard Yang, David Zhang: ShadowStream: performance evaluation as a capability in production internet live streaming networks. 347-358
• 
  
  
  
  24  Xi Liu, Florin Dobrian, Henry Milner, Junchen Jiang, Vyas Sekar, Ion Stoica, Hui Zhang: A case for a coordinated internet video control plane. 359-370
• 
  
  • 25
  Hongqiang Harry Liu, Ye Wang, Yang Richard Yang, Hao Wang, Chen Tian: Optimizing cost and performance for content multihoming. 371-382

这几篇论文都主要研究在CDN网络中由于对于视频的需求越来越大。无论是对于live的视频加以更好的性能评估还是对于如何更好的在用户与多CDN节点之间进行协调，还是如何对于分布式多个节点源的CDN网路进行最优化处理，多是重要的课题。这三篇文章就文解决这些问题提供了初步的思路。

（Ricahard Yang的论文内容确实够硬，理论十分的扎实，同时各种图表说明得到的数据十分的充分，实验做的也是十分的有说服力。）

session 9 routing

26，
Mingchen Zhao, Wenchao Zhou, Alexander J. T. Gurney, Andreas Haeberlen, Micah Sherr, Boon Thau Loo: Private and verifiable interdomain routing decisions. 383-394

介绍AS之间如何对于routing decision进行正确性验证，同时保护各自AS的policy隐私。

27，Ethan Katz-Bassett, Colin Scott, David R. Choffnes, Ítalo Cunha, Vytautas Valancius, Nick Feamster, Harsha V. Madhyastha, Thomas E. Anderson, Arvind Krishnamurthy: LIFEGUARD: practical repair of persistent route failures. 395-406

本文针对Internet中可能会出现较长时间的outage导致avalibility极具受损的问题，提出了一种新的failure定位以及修复的方法。

28，George Nychis, Chris Fallin, Thomas Moscibroda, Onur Mutlu, Srinivasan Seshan: On-chip networks from a networking perspective: congestion and scalability in many-core interconnects. 407-418

本文的背景是on-chip networks，这也是第一次接触。主要是说随着体系结构上越来越多的使用多核CPU，在多核CUP之间同样需要数据的交换以完成任务。而现在的CPU数目已经达到几十个，未来即将达到几百几千个，原先的那些集中式的busline-based、crossbar-based等方法都有缺点，所以开始考虑核与核之间的数据交换采用网络的形式，这样就使得这一架构更像网络。所以本文针对采用这种设计后可能会产生的congestion和scalibility问题进行了研究并给出了初步的解决方案。

session 10 Data centers: network resilience

29，Xin Wu, Daniel Turner, Chao-Chih Chen, David A. Maltz, Xiaowei Yang, Lihua Yuan, Ming Zhang: NetPilot: automating datacenter network failure mitigation. 419-430

本文主要考虑在datacenter DCN 网络中，当出现故障时，不是立即试图找到故障原因还是立即开始对故障进行控制即通过mitigation来保证服务仍然能够基本运行。基本的思路是试错，通过一定的方法尽量减少试错的次数，从而保证网络服务的快速恢复。

30，Peter Bodík, Ishai Menache, Mosharaf Chowdhury, Pradeepkumar Mani, David A. Maltz, Ion Stoica: Surviving failures in bandwidth-constrained datacenters. 431-442

这篇文章同样考虑如何均衡Data Center内部的健壮性tolerance和bandwidth之间的tradeoff。

31，Xia Zhou, Zengbin Zhang, Yibo Zhu, Yubo Li, Saipriya Kumar, Amin Vahdat, Ben Y. Zhao, Haitao Zheng: Mirror mirror on the ceiling: flexible wireless links for data centers. 443-454

这篇文章介绍如何更好的在DataCenter的多rack之间通过无线（60GHz）方便机架的搭建等任务，同时保障较少的受到line-of-sight等特点的干扰。