Sandstorm : even faster TCP

Researchers have worked on improving the performance of TCP since the early 1980s. At that time, many researchers considered that achieving high performance with a software-based TCP implementation was impossible. Several new transport protocols were designed at that time such as XTP. Some researchers even explored the possibility of implementing transport protocols. Hardware-based implementations are usually interesting to achieve high performance, but they are usually too inflexible for a transport protocol. In parallel with this effort, some researchers continued to believe in TCP. Dave Clark and his colleagues demonstrated in [1] that TCP stacks could be optimized to achieve high performance.

TCP implementations continued to evolve in order to achieve even higher performance. The early 2000s, with the advent of Gigabit interfaces showed a better coupling between TCP and the hardware on the network interface. Many high-speed network interfaces can compute the TCP checksum in hardware, which reduces the load on the main CPU. Furthermore, high-speed interfaces often support large segment offload. A naive implementation of a TCP/IP stack would be to send segments and acknowledgements independently. For each segment sent/received, such a stack could need to process one interrupt, a very costly operation on current hardware. Large segment offload provides an alternative by exposing to the TCP/IP stack a large segment size, up to 64 KBytes. By sending an receiving larger segments, the TCP/IP stack minimizes the cost of processing the interrupts and thus maximises its performance.

Read more...

Broadband in Europe

The European Commission has published recently an interesting survey of the deployment of broadband access technologies in Europe. The analysis is part of the Commission’s Digital Agenda that aims at enabling home users to have speeds of 30 Mbps or more. Technologies like ADSL and cable modems are widely deployed in Europe and more than 95% of European households can subscribe to fixed broadband networks.

Read more...

Multipath RTP

Multipath TCP enables communicating nodes to use different interfaces and paths to reliably exchange data. In today’s Internet, most applications use TCP and can benefit from Multipath TCP. However, multimedia applications often use the Real Time Transport (RTP) on top of UDP. A few years after the initial work on Multipath TCP, researchers at Aalto university analyzed how RTP to be extended to support multiple path. Thanks to their work, there is now a backward compatible extension of RTP that can be used on multihomed hosts. This extension will be important to access mobile streaming websites that use the Real Time Streaming Protocol (RTSP).

Read more...

The Software Defined Operator

Network operators are reconsidering the architecture of their networks to better address the quickly evolving traffic and connectivity requirements. DT is one of them and in a recent presentation at the Bell Labs Open Days in Antwerp, Axel Clauberg gave his vision of the next generation ISP network. This is not the first presentation that DT employees give on their TerraStream vision for future networks. However, there are some points that are worth being noted.

Read more...

Lessons learned from SDN experiments and deployments

The scientific literature is full of papers that propose a new technique that (sometimes only slightly) improves the state of the art and evaluate its performance (by means of mathematical models, simulations are rarely experiments with real systems). During the last years, Software Defined Networking has seen a growing interest in both the scientific community and among vendors. Initially proposed as Stanford University, Software Defined Networking, aims at changing how networks are managed and operated. Today’s networks are composed of off-the-shelf devices that support standardized protocols with proprietary software and hardware implementations. Networked devices implement the data plane to forwarding packet and the control plane to correctly compute their forwarding table. Both planes are today implemented directly on the devices.

Software Defined Networking proposes to completely change how networks are built and managed. Networked devices still implement the data plane in hardware, but this data plane, or more precisely the forwarding table that controls its operation, is exposed through a simple API to software defined by the network operator to manage the network. This software runs on a controller and controls the update of the forwarding tables and the creation/removal of flows through the network according to policies defined by the network operator. Many papers have already been written on Software Defined Networking and entire workshops are already dedicated to this field.

A recently published paper, Maturing of OpenFlow and software-defined networking through deployments, written by M. Kobayashi and his colleagues analyzes Software Defined Networking from a different angle. This paper does not present a new contribution. Instead, it takes on step back and discusses the lessons that the networking group at Stanford have learned from designing, using and experimenting with the first Software Defined Networks that are used by real users. The paper discusses many of the projects carried out at Stanford in different phases, from the small lab experiments to international wide-area networks and using SDN for production traffic. For each phase, and this is probably the most interesting part of the paper, the authors highlight several of the lessons that they have learned from these deployments. Several of these lessons are worth being highlighted :

  • the size of the forwarding table on Openflow switches matters
  • the embedded CPU on networking devices is a barreer to innovation
  • virtualization and slicing and important when deployments are considered
  • the interactions between Openflow and existing protocols such as STP can cause problems. Still, it is unlikely that existing control plane protocols will disappear soon.

This paper is a must-read for researchers working on Software Defined Networks because it provides informations that are rarely discussed in scientific papers. Furthermore, it shows that eating your own dog food, i.e. really implementing and using the solutions that we propose in out papers is useful and has value.

Bibliography

[1] Masayoshi Kobayashi, Srini Seetharaman, Guru Parulkar, Guido Appenzeller, Joseph Little, Johan van Reijendam, Paul Weissmann, Nick McKeown, Maturing of OpenFlow and software-defined networking through deployments, Computer Networks, Available online 18 November 2013, ISSN 1389-1286, http://dx.doi.org/10.1016/j.bjp.2013.10.011.

Another type of attack on Multipath TCP ?

In a recent paper presented at Hotnets, M. Zubair Shafiq and his colleagues discuss a new type of “attack” on Multipath TCP.

When the paper was announced on the Multipath TCP mailing list, I was somewhat concerned by the title. However, after having read it in details, I do not consider the inference “attack” discussed in this paper as a concern. The paper explains that thanks to Multipath TCP, it is possible for an operator to infer about the performance of “another operator” by observing the Multipath TCP packets that pass through its own network. The “attack” is discussed in the paper and some measurements are carried out in the lab to show that it is possible to infer some characteristics about the performance of the other network.

After having read the paper, I don’t think that the problem is severe and should be classified as an “attack”. First, if I want to test the performance of TCP in my competitor’s network, I can easily subscribe to this network, in particular for wireless networks that would likely benefit from Multipath TCP. There are even public measurements facilities that collect measurement data, see SamKnows, the FCC measurement app, speedtest or MLab.

More fundamentally, if an operator observes one subflow of a Multipath TCP connection, it cannot easily determine how many subflows are used in this Multipath TCP connection and what are the endpoints of these subflows. Without this information, it becomes more difficult to infer TCP performance in another specific network.

The technique proposed in the paper mainly considers the measurement throughput on each subflow as a time series whose evolution needs to be predicted. A passive measurement device could get more accurate predictions by looking at the packets that are exchanged, in particular the DATA level sequence number and acknowledgements. There is plenty of room to improve the inference technique described in this paper. Once Multipath TCP gets widely deployed and used for many applications, it might be possible to extend the technique to learn more about the performance of TCP in the global Internet.

Computer Networking : starting from the principles

In 2009, I took my first sabbatical and decided to spend a large fraction of my time to write the open-source Computer Networking : Principles, Protocols and Practice ebook. This ebook was well received by the community and it received a 20,000$ award from the Saylor foundation that published it on iTunes.

There are two approaches to teach standard computer networking classes. Most textbooks are structured on the basis of the OSI or TCP/IP layered reference models. The most popular organisation is the bottom-up approach. Students start by learning about the physical layer, then move to datalink, … This approach is used by Computer Networks among others. Almost a decade ago, Kurose and Ross took the opposite approach and started from the application layer. I liked this approach and have adopted a similar one for the first edition of Computer Networking : Principles, Protocols and Practice.

However, after a few years of experience of using the textbook with students and discussions with several colleagues who were using parts of the text, I’ve decided to revise it. This is a major revision that will include the following major modifications.

  • the second edition of the ebook will use an hybrid approach. One half of the ebook will be devoted to the key principles that any CS student must know. To explain these principles, the ebook will start from the physical layer and go up to the application. The main objective of this first part is to give the students a broad picture of the operation of computer networks without entering into any protocol detail. Several existing books discuss this briefly in their introduction or first chapter, but one chapter is not sufficient to grasp this entirely.

  • the second edition will discuss at least two different protocols in each layer to allow the students to compare different designs.

    • the application layer will continue to cover DNS, HTTP but will also include different types of remote procedure calls
    • the transport layer will continue to explain UDP and TCP, but will also cover SCTP. SCTP is cleaner than TCP and provides a different design for the students.
    • the network layer will continue to cover the data and control planes. In the control plane, RIP, OSPF and BGP remain, except that iBGP will probably not be covered due to time constraints. Concerning the data plane, given the same time constraints, we can only cover two protocols. The first edition covered IPv4 and IPv6. The second edition will cover IPv6 and MPLS. Describing MPLS (the basics, not all details about LDP and RSVP-TE, more on this in a few weeks) is important to show a different design than IP to the students. Once this choice has been made, one needs to select between IPv4 and IPv6. Covering both protocols is a waste of student’s time and the second edition will only discuss IPv6. A this point, it appears that IPv6 is more future-proof than IPv4. The description of IPv4 can still be found in the first edition of the ebook.
    • the datalink layer will continue to cover Ethernet and WiFi. Zigbee or other techniques could appear in future minor revisions
  • Practice remains an important skill that networking students need to learn. The second edition will include IPv6 labs built on top of netkit to allow the students to learn how to perform basic network configuration and management tasks on Linux.

The second edition of the book will be tested by the students who follow INGI2141 at UCL. The source code is available from https://github.com/obonaventure/cnp3/ and drafts will be posted on http://cnp3bis.info.ucl.ac.be/ every Wednesday during this semester.

Sources of networking information

Students who start their Master thesis in networking have sometimes difficulties in locating scientific information which is related to their Master thesis’ topic. Many of them start by googling with a few keywords and find random documents and wikipedia pages. To aid them, I list below some relevant sources of scientific information about networking in general. The list is far from complete and biased by my own research interests which do not cover the entire networking domain.

Digital Libraries

During the last decade, publishers of scientific journals and conference organizers have created large digital libraries that are accessible through a web portal. Many of them are protected by a paywall that provides full access only to paid subscribers, but many universities have (costly) subscriptions to (some of) these librairies. Most of these digital librairies provide access to table of contents and abstracts.

Magazines

Conferences

Journals

Standardisation bodies

Is your network ready for iOS7 and Multipath TCP ?

During the last days, millions of users have installed iOS7 on their iphones and ipad. Estimates published by The Guardian reveal that more than one third of the users have already upgraded their devices to support the new release. As I still don’t use a smartphone, I usually don’t check these new software releases. From a networking viewpoint, this iOS update is different because it is the first step towards a wide deployment of Multipath TCP [RFC 6824]. Until now, Multipath TCP has mainly been used by researchers. With iOS7, the situation changes since millions of devices are capable of using Multipath TCP.

From a networking viewpoint, the deployment of Multipath TCP is an important change that will affect many network operators. In the 20th century, networks were only composed of routers and switches. These devices are completely transparent to TCP and never change any field of the TCP header or payload. Today’s networks, mainly enterprise and cellular networks are much more complex. They include various types of middleboxes that process the IP header but also analyze the TCP headers and payload and sometimes modify them for various. Michio Honda an his colleagues presented at IMC2011 a paper that reveals the impact of these middleboxes on TCP and its extensibility. In a nutshell, this paper revealed the following behaviors :

  • some middleboxes drop TCP options that they do not understand
  • some middleboxes replace TCP options by dummy options
  • some middleboxes change fields of the TCP header (source and destination ports for NAT, but also sequence/acknowledgement numbers, window fields, …)
  • some middleboxes inspect the payload of TCP segments, reject out-of-sequence segments and sometimes modify the TCP payload (e.g. ALG for ftp on NAT)

These results had a huge influence on the design of Multipath TCP that includes various mechanisms that enable it to work around most of these middleboxes and fallback to regular TCP in case of problems (e.g. payload modifications) to preserve connectivity.

Of course, Multipath TCP will achieve the best performance when running in a network which is fully transparent and does not include middleboxes that interfere with it. Network operators might have difficulties to check the possible interference between their devices and TCP extensions like Multipath TCP. While implementing Multipath TCP in the Linux kernel, we spent a lot of time understanding the interference caused by our standard firewall that randomizes TCP sequence numbers.

To support network operators who want to check the transparency of their network, we have recently released a new open-source software called tracebox. tracebox is described in a forthcoming paper that will be presented at IMC2013.

In a nutshell, tracebox can be considered as an extension to traceroute. Like traceroute, it allows to discover devices in a network. However, while traceroute only detects IP routers, tracebox is able to detect any type of middlebox that modify some fields of the network or transport header. tracebox can be used as a command-line tool but also includes a scripting language that allows operators to develop more complex tests.

For example, tracebox can be used to verify that a path is transparent for Multipath TCP as shown below

# tracebox -n -p IP/TCP/MSS/MPCAPABLE/WSCALE bahn.de
tracebox to 81.200.198.6 (bahn.de): 64 hops max
1: 130.104.228.126 IP::CheckSum
2: 130.104.254.229 IP::TTL IP::CheckSum
3: 193.191.3.85 IP::TTL IP::CheckSum
4: 193.191.16.21 IP::TTL IP::CheckSum
5: 195.69.144.123 IP::TTL IP::CheckSum
6: 145.254.5.158 IP::TTL IP::CheckSum
7: 88.79.13.62 IP::TTL IP::CheckSum
8: 81.200.194.234 IP::TTL IP::CheckSum
9: 81.200.197.9 IP::TTL IP::CheckSum
10: 81.200.198.6 TCP::CheckSum IP::TTL IP::CheckSum TCPOptionMaxSegSize::MaxSegSize -TCPOptionMPTCPCapable -TCPOptionWindowScale

At each hop, tracebox verifies which fields of the IP/TCP headers have been modified. In the trace above, tracebox sends a SYN TCP segment on port 80 that contains MSS, MP_CAPABLE and WSCALE option. The last hop corresponds to a middlebox that changes the MSS option and removes the MP_CAPABLE and WSCALE option. Thanks to the flexibility of tracebox, it is possible to use it to detect almost any type of middlebox interference.

You can use it on Linux and MacOS to verify whether the network that you use is fully transparent to TCP. If not, tracebox will point you to the offending middlebox.