Hogging the Trough: The EFF Strikes Back

Original URL: https://www.theregister.com/2008/01/23/benett_eff_neutrality_response/

The art of voodoo network management

Posted in Networks, 23rd January 2008 17:15 GMT

The FCC is to investigate Comcast's network management practices. Last month here I gave an expert view on how the EFF, and other campaigners who called for an inquiry, don't understand the problems. Now Peter Eckserley, a copyright academic at the EFF, has responded to my article. Let's recap the story first.

To avoid congestion on its cable network, Comcast uses a technique to throttle uploads for BitTorrent users; BitTorrent downloads continue to proceed smoothly. Net Neutrality campaigners leapt onto the issue, insisting that Comcast's methods are illegitimate.

Comcast has little choice but to do what they're doing, given the three problems BitTorrent causes for their network - and their customers who don't use BitTorrent.

The first is the nature of BitTorrent itself. BitTorrent's behavior on the Comcast network is like a glutton at an all-you-can-eat buffet who insists on planting himself at the trough, preventing others from getting to the food. This causes multiple problems for DOCSIS cable networks, which caused Comcast's network managers to throttle uploads under high-load conditions (but not prohibit them outright) using a technique called Reset Spoofing.

The EFF has a preferred way of dealing with this, random packet drop. For EFF this is the One True Method of traffic management. But as I've explained to Eckersley both in print and over the phone, the problems BitTorrent causes can't be healed by random packet drop.

Packet drop would work with the regular diner who takes a plateful and moves on, but not with this super-hungry dude.

(I don't attribute malicious intent to BitTorrent's designer Bram Cohen; software often has bugs, even when it wasn't born in Redmond.)

But the EFF soldiers on, its honor now at stake.

Hogging the pipe

The EFF simply dismisses the Denial-of-Service-like effects of BitTorrent handshakes, in which a given Comcast customer's PC can become quite attractive to BitTorrent downloaders because of its favored position in the Tracker's list.

Eckersley correctly noticed that I gave an incorrect reason for this positioning: it's not because of response time, as I said, it's because of the Comcast-resident system's possession of rare file parts, which is even worse. Watch BitTorrent in operation and you'll see cycles of high popularity come and go. It's immune to packet dropping; a connect request comes into the Comcast network, and the user's system responds immediately, regardless of congestion, TCP window sizes, or load.

This highlights a design flaw in the Internet's reliance on TCP packet drop to control congestion generally; packet drop only slows traffic on established streams, not on sessions in the process of becoming established or on non-TCP streams.

Another, more persistent problem that the EFF dodges has been explained by Professor Jim Martin of Clemson University, the world's leading expert on TCP/BitTorrent interaction, and it's simple enough that a copyright expert can certainly grasp it if he wants to.

Residential networks, Comcast's being no exception, are designed on the assumption that users do more downloading than uploading. BitTorrent strives for a symmetric interchange of data, offering as much (or slightly more) in the upload direction as in the download direction. Hence, a small number of BitTorrent sessions will exhaust the network's upload capacity long before it's stressed in the download path. Professor Martin's paper, Assessing the Impact of BitTorrent on DOCSIS Networks [PDF, 450kb] predicts that fifteen BitTorrent sessions significantly slow down web browsing for the neighbors.

The web response time statistic increased from a value of 0.25 seconds when no BitTorrent users were active to 0.65 seconds when 15 BitTorrent users were active. This suggests that 15 BitTorrent users can cause a drop in performance by a factor of 2.5. When the number of BitTorrent users exceeds 30 performance degrades beyond the 1 second metric threshold.

Eckersley is certainly familiar with Martin's work, as he cites him in his original "research" on BitTorrent and Comcast.

An additional problem arises from BitTorrent's tendency to punish users on fast connections with greater traffic loads. Even though Comcast limits upstream traffic to 384kbit/s, a small fraction of its basic 4Mbit/s download rate, upstream traffic moves considerably faster on its network than it does on a standard DSL connection.

Consequently, BitTorrent downloaders will gravitate to peers on Comcast over those on DSL through their own performance assessment.

So let's recap the argument. I say BitTorrent causes problems for Comcast's DOCSIS network that can't be healed economically by randomly dropping packets. Packet drop applies back-pressure to conventional TCP sessions, but it does nothing to the data queued inside Comcast users' cable modems, where it contends for scarce upstream bandwidth with other users.

Packet-drop also does nothing to affect the rate at which new connect requests come into the Comcast network from BitTorrent users across the Internet. Random packet-drop has profound effects on conventional applications, but very little on BitTorrent, and the routers to do packet-drop in real-time are more expensive than the aynchronous Sandvine system.

Therefore, it's acceptable for Comcast, as a matter of reasonable network management, to employ TCP Resets to prevent BitTorrent doing harm to the web browsing, standard file downloading, and VoIP sessions that are the typical behavior of the Comcast customer.

My claim is borne out by Professor Martin's analysis of TCP's interaction with the DOCSIS Medium Access Control protocol, and by his analysis of BitTorrent's side effects on web browsing. It's further borne out by observations of BitTorrent's cycling behavior, and the well-known weakness in packet drop in terms of cycling and fairness which I addressed on the last article.

Confronted with all this evidence, the EFF and Eckersley would be wise to admit their confusion and vow to devote further study to the topic, in particular to wait for the FCC's analysis of the complaint.

To the cable company's credit, it has announced plans to begin a major network upgrade starting in the fourth quarter, resulting in a nearly symmetric 130/100 Mbit/s network with the kind of traffic controls that the EFF dreams about.

(When it's not lobbying for usage-based pricing. Eckersley cites Australian ISP Whirlpool's pricing plans as exemplary.)

With such a system on the drawing board, it's not hard to fathom why Comcast doesn't want to lay in the collection of the expensive routers (Sandvine runs on PC hardware and processes packets out-of-band) the EFF would have them buy to patch their current network into EFF-approved shape; all it takes is a willingness to see reality as it is and an ability to put the holy books aside. That may be a hard slog for a Net-Utopian advocacy group harboring delusions of conspiratorial persecution, but it's the right thing to do.

Conclusion

Everyone who's argued with religious fanatics has seen them dig in their heels and flail when confronted with challenges to their belief systems. Point out the inconsistencies in the Genesis account of creation or the implausibility of Noah's Ark and you'll get some creative sputtering followed by affirmations of faith at a high volume level.

The EFF's response is over-heated religious rhetoric that ignores both the economic constraints that bind network operators, and the ineffectiveness of the Internet's inherited methods at dealing with challenges created by new protocols and applications. The literature on packet-drop in particular suggests a dozen variations, none of which applies across the scope of a single BitTorrent user's communications.

While the Internet's end-to-end architecture makes it a fertile testbed for the implementation of new protocols, each brings with it new traffic patterns that must be dealt with in ways that prevent others from starving. Hogging the trough is simply not acceptable network behavior.®

Richard Bennett is a network architect and occasional activist in Silicon Valley. He wrote the first standard for Ethernet over twisted-pair wiring and contributed to the standards for WiFi and the Ultra-Wideband wireless networks. His eleven-year old blog is at bennett.com. He will debate these issues on a panel at the Toll Roads Symposium in San Francisco on Saturday.