Congestion Control Chapter 6

1. Congestion ControlChapter 6 Outline Resource Allocation Issues Queuing Disciplines FCFS (FIFO queues) Priority Queuing Fair Queuing (for flows) TCP Congestion Control Detection – Resolution approach (AIMD and Slow Start) Alternatives: Fast Transmit / Fast Recovery Congestion Avoidance router-centric: DECbit and RED Gateways host-centric: TCP Vegas QoS

2. Source 1 10-Mbps Ethernet Router Destination 1.5-Mbps T1 link 100-Mbps FDDI Source 2 Congestion Control ISSUES: • How to fairly allocate resources (link bandwidths and switch buffers) among users. • Two sides of the same coin: • Resource allocation so as to avoid congestion (difficult with any precision) • Congestion control if (and when) it occurs • Resource allocation and congestion control involve both: • hosts at the edges of the network (transport protocols) • routers inside the network (queuing disciplines) • Underlying service model can be • best-effort (assume here – end-hosts given no opportunity for QoS demands) • multiple qualities of service QoS (later) Congestion in a packet-switched network

3. Source 1 Router Destination 1 Router Source 2 Router Destination 2 Source 3 Framework Multiple flows passing thru a set of routers • Connectionless flows assumed: What are they? Even tho datagrams from a source to a dest are switched independently, they typically flow thru the same path. • Routers maintain soft state info • Somewhere between the hard state info of a VC switch (bandwidth, cell-loss ratio, etc) and no state info of pure connectionless. • Correct operation does not depend on soft state info but is improved by it. • Implicitly defined: router watches for what appears to be a flow – used in TCP Congestion Control. • Explicitly defined: source sends flow-setup (flow about to start) across network. (a step down from a VC since explicit flow has no reliable, ordered delivery) • Taxonomy of Resource Allocation/Congestion Controlmechanisms • Router-centric: address prob inside net (decide forwards/drops, inform hosts) versus Host-centric: address problem from outside the network) • Reservation-based: hosts request capacity when flow is established; versus Feedback-based:Explicit(e.g., congested router sends “slow-down message) Implicit(eg, host adjust rate based on, eg, cell-loss rate) • Window-based (telling sender remaining buffer space – as in flow control) versus Rate-based (telling sender the rate at which data can be absorbed)

4. Throughput/delay Optimal Load load Evaluation Criteria(of resource allocation effectiveness & fairness) Thrashing or congestion collapse Power • Effective Resource Allocation (utilization issue – network-wide point of view) measured by Power = ratio of thruput to delay. • Fair Resource Allocation (to individual senders) • Can assume Fair means Equal shares • E.g., Raj Jain proposed metric when Fair means Equal and all paths are equal length: Jain’s Fairness Index: Given flow thruputs (units/sec) x1, x2, …, xn f(x1, x2, …, xn) = ( ni=1 xi)2/(n ni=1 xi2 ) If all n flows have thruput of 1 unit/sec, f = n2 / n*n = 1. However if k have thruput 1 and n-k have thruput 0, f = k2 / n*k = k/n (less fair)

5. Flow 1 Flow 2 Round-robin service Flow 3 Flow 4 Queuing Discipline (Each router specifies a queuing discipline regardless of resource allocation mechanism. Algorithm can be thought of as allocating bandwidth (which packets get transmitted) and buffer space (which packets get dropped)) • First-In-First-Out or FIFO (AKA: FCFS) • Packets transmitted in arrival order. • No discrimination between traffic sources. • Usually used with “tail drop” policy. • FIFO + tail-drop = bundle. • Widely used in Internet. • Variations include priority queuing. • Fair Queuing (FQ) for Flows • explicitly segregates traffic based on flows (separate queue per flow) • Weighted Fair Queuing allows a weight to be assigned to each flow.

6. Flow 1 Flow 2 Flow 1 Flow 2 Output (arriving) (transmitting) Output F = 10 F = 10 F = 8 F = 5 F = 2 (a) (b) Fair Queuing - FQ Algorithm For simplicity, suppose clock ticks each time bit is transmitted (bit = tic) Let Pi = length of packet i Si = time when transmission of packet i starts Fi = time when transmission of packet i finishes Fi = Si + Pi For a single flow, when does a router start transmitting packet i? if it’s before router is finished with this flow’s packet i-1, right after last bit of i-1 (Fi-1) if no current packets for this flow, then start transmitting when 1 arrives (at time Ai) Thus: MAX (Fi - 1, Ai) and Fi = MAX (Fi - 1, Ai) + Pi For multiple flows(Not perfect: can’t preempt current packet) calculate Fi for each packet that arrives on each flow (treat as timestamps) packet with lowest timestamp is next. Queue discipline: Shortest packet first Longer packet already in progress is completed first

7. TCP Congestion Control • Idea • assumes best-effort network (FIFO or FQ routers) each source determines network capacity for itself • uses implicit feedback (host adjusts rate based on its knowledge) • ACKs pace transmission (self-clocking) (I.e., only allow n outstanding un-Ack’ed packets. • Challenge • determining the available capacity in the first place • adjusting to changes in the available capacity • AIMD and Slow Start were the original solutions for TCP

8. Additive Increase/Multiplicative Decrease (AIMD) Objective: adjust to changes in the available capacity • New state variable per connection: CongestionWindow • set by source to limit number of packets in transit • Recall, FlowCtrl AdvertisedWindow = # of packets destination can still buffer) MaxWin = MIN( CongestionWindow, AdvertisedWindow ) EffWin = MaxWin - ( LastByteSent - LastByteAcked ) # ofoutstanding packets • Idea: • increase CongestionWindow when congestion goes down • decrease CongestionWindow when congestion goes up • Question: how does the source determine whether or not the network is congested? • Answer: a packet timeout occurs (I.e., an Ack is late) • Assumes timeout signals that a packet was dropped due to congestion (packet loss is so seldom due to transmission error) • lost packet implies congestion

9. Source Destination 70 60 50 40 KB 30 20 10 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0 T ime (seconds) … Algorithm: Each time source successfully sends a CongestionWindow of packets, increase CongestionWindow by 1 packet (additive incr). Divide CongestionWindow by 2 each timeout (multiplicative decr) (never below Min Seg Size – MSS is in bytes – usually 1 packet) AIMD (cont) • In practice however, TCP increments a little for each ACK, using: Increment = MSS * (MSS/CongestionWindow) CongestionWindow += Increment Trace: CongestionWindow sawtooth behavior with AIMD AIMD works well when source is operating close to the available capacity of the network. But takes too long to ramp up from scratch. SLOW START (ironically name) is intended to solve that using multiplicative increase.

10. Source Destination 70 60 50 KB 40 30 20 10 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 … Slow Start (2nd mechanism provided by TCP) • Start with CongestionWindow (CW) = 1 packet a slow start compared to a CongestionWindow=AdvertisedWindow start • Double CongestionWindow each RTT (multiplicative incr) until it reaches CongestionThreshold (CT), then increment by 1 per RTT. Used when first starting connection and if connection goes dead waiting for timeout (another “start over” situation). Slow Start Trace: No increase; No Acks arriving – due to lost packets Timeout; 17=CTCW/2; CW  0 Multiplicative increase until CT, then Additive increase No increase; No Acks arriving Hash marks =times when each packet is transmitted timeouts time when retransmitted packets were first transmitted mult increase Time in sec Multiplicative increase until CT, then Additive increase Timeout; 11=CTCW/2; CW  0

11. Sender Receiver Packet 1 Packet 2 ACK 1 Packet 3 70 ACK 2 Packet 4 60 50 ACK 2 Packet 5 40 KB 30 Packet 6 20 ACK 2 10 ACK 2 1.0 2.0 3.0 4.0 5.0 6.0 7.0 Retransmit packet 3 ACK 6 Fast Retransmit Problem: Coarse-grain TCP timeouts lead to idle periods Fast retransmit: use duplicate ACKs to trigger retrans. Idea: every time a packet arrives, receiver sends ACK. Thus, when a packet arrives out-of-order (and TCP can’t ACK because earlier packets have not yet arrived) TCP resends last legit cumm ACK (called duplicate ACK). When sender sees 3 dups, retransmits next packet. Trace of CongestionWindow with fast retransmit Fast Recovery:Upon congestion, rather than drop back to 0 and use Slow Start, just cut window in half and resume additive increase. Hash marks =times when each packet is transmitted timeout time when retransmitted packets were first transmitted Time in sec Eliminates many of the flat areas where no packets were transmitted

12. Congestion Avoidance • TCP’s strategy is to control congestion once it happens (repeatedly increase load to find the point at which congestion occurs, and then back off) • Alternative strategy • predict when congestion is about to happen • reduce rate before packets start being discarded • call this congestion avoidance, instead of congestion control • Two possibilities • router-centric: DECbit and RED Gateways • host-centric: TCP Vegas

13. Queue length Queue length Current time T ime Previous Current cycle cycle A veraging interval DECbit • Add congestion bit to packet header. • Router • monitors average queue length over last busy-idle cycle + current busy cycle, set congestion bit if average queue length > 1 • End Host • Destination echoes bit back to source • Source records how many packets resulted in set bit • If less than 50% of last CongestionWindow’s worth had bit set • increase CongestionWindow by 1 packet • If 50% or more of last window’s worth had bit set • decrease CongestionWindow to 7/8th of its value. Random Early Dectection (RED) • Notification is implicit • Router just drops the packet when congested (TCP will timeout) • Early random drop • rather than wait for queue to become completely full, drop each arriving packet with some drop probability whenever the queue length exceeds some drop level

14. MaxThreshold MinThreshold P(drop) 1.0 A vgLen MaxP A vgLen MaxThresh MinThresh RED Details Weighted runnng avg queue length Compute average queue length AvgLen= (1-Weight)*AvgLen+Weight*SampleLen 0 < Weight < 1 (usually 0.002) SampleLen= queue length each time packet arrives Two queue length thresholds if AvgLen  MinThresholdthen enqueue packet if MinThreshold < AvgLen < MaxThreshold then calculate probability P drop arriving packet with probability P if MaxThreshold  AvgLen, then drop arriving packet Computing probability P TempP = MaxP * (AvgLen - MinThreshold) (MaxThreshold - MinThreshold) Count = # packets (denom of AvgLen) P = TempP/(1 - count * TempP) Drop probability curve

15. TCP Vegas (host-centric congestion avoidance) min of all measured RTTs, Typically RTT of 1st packet Idea: source watches for some sign router’s queue is building (eg, RTT grows; sending rate flattens) ExpectedRate =CW/BaseRTT Diff = ExpectedRate – ActualRate if Diff < α increase CW linearly else if Diff > β decrease CW linearly else leave CW unchanged ( when α < Diff < β ) Source calculates current sending rate as the # bytes divided by the RTT for a distinguished packet roughly corresponds to too little data in the network roughly corresponds to too much data in the network

16. TCP Vegas(trace of congestion avoidance mechanism) Congestion Window Trace for TCP Vegas Parameters a = 1 packet b = 3 packets Actual throughput Expccted throughput Shaded area is region between a and b units away From the Expected throughput (the goal to keep actual in this region. Note the actual gets drug along by shaded.)

17. Sampler , Microphone Buffer , A D D A converter Speaker QoSReal-time App • Require “deliver on time” assurances • must come from inside the network (hosts cannot make such guarantees alone) • Example application (audio) • sample voice once every 125us • each sample has a playback time • packets experience variable delay in network • add constant factor to playback time: playback point • Playback Buffer Packet generation Packet arrival Buffer Sequence number Network delay Playback Time

18. Integrated Services • Refers to the body of work by IETF 1995-97 working group on Integrated Services. • Integrated Services allocates resources to individual flows • whereas Differentiated Services allocates resources by “classes of traffic” • Integrated Service Service Classes • E.g., Guaranteed service (packets are never late – guaranteed max delay time) • Flowspecs (Set of info we provide to the network to specify needs.) • Tspec • describes flow’s Traffic characteristics (e.g., average bandwidth, token issues..) • Rspec • describes the services Requested from the network • E.g., guarantees, such as, delay target

19. RSVP Resource reSerVation Protocol • While connection-oriented networks have setup protocols, best-effort connectionless networks don’t – they need some sort of reservation protocl in order to offer QoS. • Internet resource reservation corresponds to signaling in ATM • Proposed Internet standard is called RSVP • Receiver-oriented • 2 messages: PATH and RESV • Source transmits PATH messages every 30 seconds to make requests. • Destination responds with RESV message to ack.

20. RSVP versus ATM (Q.2931) • RSVP • receiver generates reservation • soft state info used in routeers (it is refreshed/timedout) • separate from route establishment • QoS can change dynamically • ATM • sender generates connection request • hard state info (requires explicit delete at teardown) • concurrent with route establishment • QoS is static for life of connection

21. Differentiated Services (also IETF) • Problem with Integrated Services: scalability • Idea of Differentiated Serivces: support 2 classes of packets • DS adds new Premium Service to best effort traffic class) • Which packets are premium? • Use premium-bit in header.