Congestion and its control: Broadband access networks. David Clark MIT CFP October 2008

Transcription

1 Congestion and its control: Broadband access networks David Clark MIT CFP October 2008

2 Outline Quick review/definition of congestion. Quick review of historical context. Quick review of access architecture. Discussion of possible approaches I hope to augment this during this workshop.

3 The picture I will use this as a reference: Switch, concentrator, etc. Incoming links, with total offered load L A queue of packets waiting for service Link, of capacity C

4 The words If more traffic L arrives than can be carried over the link with capacity C, packets pile up in the queue. If the queue fills up, packets get dropped. If this situation persists, something better happen to throttle the incoming traffic or the system becomes unstable.

5 Definitions of congestion Formal: Congestion = L/C Or if you like Greek or are a queuing theorist: ρ = λ / µ Networking: Sometimes the same, sometimes when the queue overflows.

6 Definitions, continued Econonomics: When an increase in the use of a facility or service which is used by a number of people would impose a cost (not necessarily a monetary cost) on the existing users, that facility is said to be congested. (Thanks to Steve Bauer for tracking this down.) Sort of similar, but not related to mechanism. Cost might be reduced throughput or increased delay. If only one user overloads the link, not congestion, just overload.

7 Definitions, continued Network operator: [O]ver any 15-minute period, if an average of more than 70 percent of a port s upstream bandwidth capacity or more than 80 percent of a port s downstream bandwidth capacity is utilized, that port will be determined to be in a Near Congestion State. From Comcast description of current management practices.

8 Consider this last definition Congestion (or near-congestion) is described as behavior over some period of time. Not instantaneous. This longer-term measure may be a reasonable proxy for congestion, but is only an approximation. System can run at 99% loading and never generate perceived congestion if load is non-varying. System can operate at 50% loads and have moments of intense congestion that are objectionable to users. Return later to why this measure is useful.

9 Marketing congestion (Not a term in common use, but useful to have.) When there is enough traffic that the provider cannot meet the committed rates to its customers. Users may be rate limited even when the system is underloaded, for marketing reasons. So they will encounter congestion with underloaded links. ISPs have to play the statistics games, and sometimes it leads to overloads. Marketing congestion has little to do with actual congestion, but relates to meeting expectations.

10 History of congestion In the beginning, we had no controls, and Internet was in congestive collapse. Around 1986, Van Jacobson proposed a scheme (called slow-start ) that got congestion under control. It has been elaborated and enhanced, but we still use it. Defined as part of TCP, but other protocols are expected to conform. The TCP-friendly requirement.

11 Key features TCP keeps speeding up until it gets a slowdown signal. It then (more or less) cuts its sending rate in half and repeats. Produces a saw-tooth sending rate. The slow-down signal can be a dropped packet or an explicit congestion notification (ECN) signal. Implemented in the sending end-node. We trust the end-node to do the right thing.

12 Implications TCP will just keep speeding up until: It gets a slow-down signal In other words, there is congestion. It runs out of data. It tops out either end-point processor. Modern PC can get close to a gb/s today. So there will always be congestion unless there is nothing to send. The economist might not always call this congestion, but the queuing theorist would.

13 Sub-definition Self-congestion When one user generates enough traffic to overload a link. Shared congestion When multiple uses combine to overload (congest) a link.

14 If congestion always occurs Where? In the old days of dialup, the modem. Now, less clear. Might be the access link. Might be inside the network. For example, at a peering point. Might be at the server. But there will be a bottleneck somewhere. Unless there is nothing to send.

15 Symptoms of congestion Limits on rate. As implemented, TCP hunts (with that saw-tooth pattern) for a rate that will not congest the bottleneck link on average. Variation in rate. That saw-tooth. More or less a factor of two. Variation (and increase) in delay. Term for variation is jitter. Because packets sit in that queue before being sent.

16 Queues cause latency Switch, concentrator, etc. Incoming links, with total offered load L A queue of packets waiting for service Link, of capacity C

17 Which one matters? Users measure speed. It is how their service was characterized. Bulk throughput is easy to grasp. But for web browsing, it often is not the factor in user experience. Many applications (and thus many users) are very sensitive to latency. VoIP, games. Gamers know all about ping times, latency, and lag.

18 Queues--the changing practice The old theory: put enough buffering to hold one roundtrip worth of data. Cross-country in Internet today is 100 ms. So add 100 ms of buffering. It is more useful to measure buffers in terms of delay, not bytes. But the conversion is easy The new idea: in the core of the network, you don t need nearly this much. A small number of packets. So the real fast,expensive routers don t need a lot of expensive buffers. But near the edge, the round trip rule may still hold. Consequence: see latencies between 100 and 200 ms.

19 Modern congestion controls To control jitter (variation in latency): Add more sophisticated algorithms at point of congestion to send congestion signal when queue starts to form, not when it overflows. Advanced schemes are called Active Queue Management (AQM). Measure variation in latency at the sender, and try to adapt rate so as to control variation. This (I believe) is what BitTorrent DNA does.

20 BitTorrent BitTorrent DNA is able to automatically detect game play on a user s system and through BitTorrent s advanced proprietary congestion control technology eliminates any impact on game play as well as other sensitive real-time applications including Voice over IP (VoIP) calls, streaming media, or Internet browsing. BitTorrent Press Release, Sept 15, 2008.

21 HFC--a case study The home: upstream traffic Home router Coax In-home links to PCs Cable modem

22 Upstream congestion Occurs for two reasons: Shared congestion on the HFC system. Enough users show up at once so that system cannot meet its commitments. This is example of what I called marketing congestion. Self congestion: Combination of traffic from one home overloads the upstream commitment. Self-congestion from the perspective of the ISP. Shared congestion from the perspective of the individual users in the home. Both produce the same latency consequences.

23 Dealing with (shared) congestion Let TCP do its thing Flows with shorter round trips go faster. Is this what you want? Users with more flows go faster. Since TCP separately limits each flow. Rate limit some aggregate E.g. the flows of a customer. What Comcast is currently proposing. Give some flows priority at point of congestion. Give some flows a different share. Both of these could be based on many criteria.

24 Priority as a mechanism What it does: By allowing packets from certain flows to jump to the head of the queue: Limited variation in latency (great for games and VoIP. Allows these flows to grab all the bandwidth, if they are designed to do so. VoIP and games are not bandwidth hogs. If abused. A bandwidth-greedy app will squeeze out others. If everything is marked, it is as if nothing is marked. How to implement it. Look at QoS bits. Look at port number (well known ports-> applications)

25 Shares as a mechanism Put different flows into shares, and rate limit each share. One share cannot mess up another. Good isolation. In contrast to priority. As a share gets congested, all the same issues we have discussed occur inside each share (e.g. jitter).

26 Many papers Have been written on the interplay of priority, shares, admission control, and other schemes to deal with congestion, in the context of different operating requirements. One of the most over-studied areas of networking. In this talk, I will not dwell further.

27 Trust Should the ISP trust the user to use the QoS marking in a reasonable way? Answer: It all depends on where the queue is. If the queue is in the cable modem, then the user can only hurt himself. Self-congestion from the perspective of the ISP. If the point of congestion is deep in the network, then the marks would allow different users to compete with each other inside the network. Perhaps less sensible

28 Action at a distance As described so far, these controls (rate limits, shares or priority) have to be applied at the point of congestion. Could be in the cable modem, or deep in the network. TCP feedback allows the source to adapt to congestion that is encountered elsewhere To achieve remote control of congestion, need some sort of marking or feedback. BT DNS can be applied at the source because it can measure the round trip delay there. Other goals (such as limiting groups of flows), cannot be done at the source using TCP feedback as defined.