The Future of TCP: Train-wreck or Evolution?

Stanford Clean Slate Internet ProgramStanford University


April 1, 2008
Stanford University, Stanford, CA

Sponsored by:

The Stanford Clean Slate Internet Program
Cisco Systems
The Stanford Computer Forum


All demos will be held in Packard 101, with the exception of the dinner on 31 March as noted below.
All demos and talks will be video-taped. The videos will be available at this site after the workshop.

Monday, 31 March, 2008:

7:00 PM Dinner at Zibibbo
430 Kipling Street, Palo Alto, CA 94301

Tuesday, 1 April, 2008:

8:00 - 8:30 AM Breakfast
8:30 - 8:45 AM Introductory remarks
Nick McKeown
8:45 - 9:15 AM Demo I:
Incompatibility of TCP Congestion Control Mechanisms with State of the Art Wireless Schedulers

Ulas C. Kozat

In this demo we will focus on the inter-play of TCP congestion control mechanisms and opportunistic wireless schedulers such as Proportional Fair Sharing (PFS). Specifically, we consider a cellular setting that is capable of packet scheduling both in the forward and reverse links. Our demo will stress on the facts that (i) TCP suffers substantially when there is a high degree of rate/load asymmetry in the forward and reverse directions, (ii) TCP is vulnerable when a mobile host has payload traffic in both directions, and (iii) the fairness notion adopted at the wireless schedulers are not necessarily compatible with the per flow fairness of TCP. Since majority of video streaming applications in the current Internet use TCP, our demo will compare video streaming sessions over TCP vs. a backlogged system under the same network and wireless channel conditions.

Slides [.ppt]

9:15 - 9:45 AM Demo II:
TCP over multi-hop wireless networks

Konstantinos Psounis, Ramesh Govindan

TCP has not been designed to cope with the complex interference found in static multi-hop wireless (mesh) networks. This interference makes congestion and rate control a neighborhood, not a single-node affair, and unless TCP explicitly accounts for this, its performance can be spectacularly bad. We use simulation and experimental results to showcase this fact and explain the origins of the problem, which points to some ways of fixing it. Conceptually, such fixes are evolutionary (in the sense that one should be able to retrofit them into TCP), but one would have to change the TCP implementation a fair bit to accommodate this.

Slides [.pdf]
Demo Video I [.mov]
Demo Video II [.mov]
Demo Video III [.mov]

9:45 - 10:30 AM Invited Talk I:
TCP Congestion Control Evolution

Steven Low

It is unlikely train wreck will occur to TCP congestion control. The current mechanism has shown remarkable robustness in the face of the explosive growth of the internet, in size, in heterogeneity, in traffic, and in applications. However, I will argue that TCP evolution is a series of punctuated equilibria and that now is again a time to change, and will support my argument with real-life data.

Slides [.ppt]

10:30 - 10:45 AM Break
10:45 - 11:45 PM Demo III:
Injong Rhee, Ajit Warrier, Sangtae Ha

a) TCP performance problems in wireless multi-hop networks

This demo demonstrates the fairness problem of TCP when multiple TCP flows run in wireless multi-hop networks. The reason why TCP can ensure fairness in wired Internet that all competing flows experience the same level of congestion when they go through the same bottleneck link. However, in wireless multi-hop networks, two competing flows at the same bottleneck may see a different amount of congestion. This happens because interference is one of the causes for congestion in such networks and the level of interference that a flow experiences, depends on the locations of its source and destination nodes. The imbalance in interference and consequently congestion causes the flows with more interference to reduce their rates while the flows with less interference see the rate reduction as increase in available bandwidth and increase their rates. The vicious cycle of rate reduction and increase causes the eventual starvation for the flows with more interference.

Slides [.ppt]

b) TCP problem under high bandwidth delay product networks.

Slides [.ppt]

11:45 - 12:15 PM Demo IV:
Video Streaming Over Wireless: Where TCP is not enough

Xiaoqing Zhu, Jatinder Pal Singh, Bernd Girod

This demo highlights the limitation of TCP for supporting video streaming over wireless, in the presence of heterogeneous link speeds. The setup mimics a wireless home network with several nodes operating in 802.11 ad-hoc mode. We demonstrate the impact of a file transfer session over a slow link on an ongoing video streaming session over a fast link. TCP leads to approximately equal throughput for both sessions, despite difference in their link speed, causing severe quality degradation of the video stream.

Slides [.ppt]
Demo Video I [.asf]
Demo Video II [.asf]

12:15 - 1:30 PM Lunch (held at Packard 101)
1:30 - 2:15 PM Invited Talk II:
TCP Issues in the Data Center

Tom Lyon, Nuova Systems

The number and size of datacenters are growing exponentially, driven by businesses such as Google, Amazon, and the Wall Street trading firms. Essentially all the network applications in these datacenters depend on TCP. Yet in datacenters, the bandwidth, latency, cost, and congestion parameters that affect TCP are drastically different than in the wide-area Internet. We explore a few issues with TCP and demonstrate that its implementation today is far from optimal for the datacenter.

Slides [.ppt]

2:15 - 2:45 PM Demo V:
TCP for home networking: Why you can't teach an old dog new tricks

Hariharan Rahul, Szymon Chachulski, Kah Keng Tay, and Dina Katabi

Multimedia home networks are becoming big business, with over 30 million households expected to have multimedia networks by 2010, and home entertainment projected to reach $12B by 2009. Users can use home entertainment applications, many of which currently run on TCP, to stream audio and video to many rooms, share photos, songs and movies, and support security cameras and data applications. In this talk, we will demonstrate that TCP's architecture, which was designed for elastic applications on wired networks, is fundamentally not suited either to the challenges posed by lossy wireless networks and streaming multicast applications, or to the opportunities presented by wireless mesh networks.

Slides [.pptx]
Demo Video I [.mpeg]
Demo Video II [.mpeg]
Demo Video III [.mpeg]
Demo Video IV [.mpeg]

2:45 - 3:30 PM Invited Talk III:
In defense of TCP

Balaji Prabhakar, Stanford University

As TCP turns 20, it can be proud of enabling the Internet phenomenon. Recently, it has been found to be wanting in several ways; notably, in its performance in large bandwidth delay product networks, being sluggish, needing large buffers, and its treatment of short flows. We find that some of these shortcomings are either due to unreasonable expectations, or are easily addressable by small changes to TCP; i.e. that TCP is fundamentally sound, it could use a "facelift."

Slides [.ppt]

3:30 - 3:45 PM Break
3:45 - 4:15 PM Demo VI:
Congestion Collapse in Grid 5000

Romaric Guillier, Pascale Vicat-Blanc Primet

Today, the capacity and the usage of the Internet is fundamentally changing. In the forthcoming year, millions of homes will have access to the Internet with Fiber lines. Network companies will offer speeds on Fiber of up to 1Gbps. Consequently, ultra-high-speed applications will be enabled at low-cost, such as high-definition teleconferencing, telemedicine and advanced telecommuting for people working from home. We then propose a demonstration which aims at showing that the traditional congestion control approach may present a strong barrier to the large scale deployment of these exciting and very useful envisioned applications exploiting the emerging huge access capacities.

Slides [.pdf]

4:15 - 4:45 PM Demo VII:
Short Messages

Damon Wischik
4:45 - 5:15 PM Demo VIII:
TCP in a world of cloud services

Jiang Zhu, Nandita Dukkipati, Sateesh Addepalli, Flavio Bonomi

Cloud services such as those provided by Amazon, Microsoft, Google are proliferating. On one hand they offer the convenience of services such as computing and storage on demand, and accessible content anytime anywhere; on the other hand, access to these services can be painfully long through WAN links which often have long round-trip times, are lossy, and shared across a lot of simultaneous users. In this demo, we take a particular example of uploading user-generated content such as HD video clips to the cloud. We are interested in the metric -- how quickly is the content shareable? The demo illustrates the painfully long "Time to Share" in part due to TCP congestion control, and contrasts it with the quick sharing time of an "ideal" solution.

Slides [.pdf] [.mov]

5:15 - 5:30 PM Concluding remarks
Nick McKeown