Outline

• Queueing Disciplines
• TCP Congestion Control
• Quiz
• TCP Congestion-Avoidance

Queueing Disciplines

• Every router must implement some queueing discipline to govern how packets are buffered while waiting to be transmitted.
• This discipline can be viewed as allocating both bandwidth (saying which packets gets to be transmitted) and allocating buffer space (which packets get discarded).
• The discipline also affects the latency experienced by the packet.
• We are now going to look at two queueing algorithms in terms of forwarding: first-in first-out (FIFO) and fair queueing (FQ).

FIFO Queueing

• When using FIFO queueing the first packet to arrive at a router is the first to be forwarded.
• It doesn't matter which flow the packet belongs to or how important the packet is.
• When the queue becomes full, then the router drops incoming packets until there is more space in the queue.
• This is called tail drop.
• Here FIFO queueing is a scheduling policy, it determines the order in which packets are transmitted and drop tail is a drop policy, it determines which packets are dropped.
• FIFO with tail drop is the most widely used policy in internet routers today.
• Notice this queueing mechanism does not take any account of congestion, so congestion mitigation has to be done on top of this.
• A variation on FIFO is priority queueing. The queue is split into priority tiers using the IP Type of Service (TOS) field, the first-in packet of the most important tier is sent next.
• The problem with priority queueing is that the highest priority queue can starve all the other queues.

Fair Queueing

• FIFO queueing doesn't differentiate between traffic sources. If priorities are being used it doesn't police how the sources are using the priorities.
• Fair queueing tries to address these issues.
• It separates the queue into distinct flows being handled by the router.
• The router then tries to handles the flows in a round-robin fashion.
• When a queue for a given flow reaches a certain length additional packets sent to that queue are discarded.

Fair Queueing Details

• To make this more precise let S_i denote when the router starts to transmit packet i, and F_i denote when it finished.
• Then P_i = F_i - S_i represents the time to send the ith packet.
• Finally, let A_i denote the time the ith pack arrives at the router.
• Then F_i = max(F_i-1, A_i) + P_i
• In fair queueing, we calculate these F_i independently for each flow.
• We choose the next packet to send as the one whose F_i is the lowest.
• Notice this queueing discipline is work-conserving. That is, it never leaves the link idle as long as there is a least one packet in the queue.
• Another thing to notice is that if the link is fully loaded and there are n flows, then each flow can use at most 1/n of the total bandwidth.
• A variation of FQ is weighted fair queueing. In this, each queue has a weight and we multiply the F_i's by this weight and choose the lowest product as who to send next.

TCP Congestion Control

• In addition to the flow control advertised window, TCP also maintains a CongestionWindow state variable.
• The equation we had for EffectiveWindow a couple lectures back is not exactly what is used in practice.
• Instead, we set
  MaxWindow = MIN(CongestionWindow, AdvertisedWindow), and
  EffectiveWindow = MaxWindow - (LastByteSent - LastByteAcked)
• Unlike the AdvertisedWindow which the receiver controls, the CongestionWindow is determined by the sender.
• The sender tries to increase the window size when the level of congestion goes down and decrease it when it goes up. The acronym for this is AIMD (additive increase.multiplicative decrease).
• Timeouts/segment losses are used as an indication of congestion. When a timeout/loss occurs, the congestion window is halved (explaining the MD in AIMD).
• The window size is not allowed to go below one MSS (maximum segment size).
• On the other hand, every time TCP succeeds in sending and getting ACK's for a segment, we increase the CongestionWindow by MSS/CongestionWindow. (This is AI in AIMD).
• The reason why we add to increase and multiply to decrease is that the consequences of having a too large window (which would affect other people's traffic) are worse than those of being too small (only your traffic).

Quiz

Which of the following is true?

1. The way TCP usually extends the 16 bit window advertise size x is to use an additional 32 bit option y, such that the window is y*2^{16} + x.
2. RTCP is the protocol that is used to send mainly data such as video when RTP is being used.
3. Within an RPC channel request reply messages are sequential.

Slow Start

• Additive increase works well, when the source is operating at close to the available capacity of the network.
• It takes too long to ramp up a connection when it is starting from scratch.
• TCP allows for a second mechanism called slow start to solve this problem.
• The name slow start comes by comparison with how the EffectiveWindow is computed without any congestion control.
• In slow start, when a connection is first set-up the CongestionWindow is set to one MSS.
• After the first ACK, it doubles to 2 MSS's, then 4, etc.
• When a packet loss occurs, the window is halved. This window size is called the CongestionThreshold.
• We then do increases based on 1 ,2 ,4, etc. MSS's. When we reach the original CongestionWindow + CongestionThreshold size we resume doing only additive increases.

Fast Retransmit and Fast Recovery

• TCP timeouts are relatively course grained. So one could end up using our scheme so far with dead-air while waiting for a timer to time out on a lost packet.
• To avoid this a mechanism called fast retransmit was added to TCP.
• When fast retansmit is being used, the receiver ACK's each packet it receives even if this is out of order. The ACK is still for the next byte expected.
• This might result is several duplicate ACKs. If the sender sees three duplicates ACKs it retransmits.
• Fast retransmit tends to work well when the window size is at least large enough so one can expect to see these three timeouts before the timer goes off.
• Another mechanism, called fast recovery can be used to avoid having to halve the CongestionWindow on packet loss.
• Rather than halving the whole congestion window. We take the average of the size that caused a problem and the Congestion Threshold.
• This allows us to binary search for the correct window size.

Congestion-Avoidance Mechanisms

• As described so far, TCP need needs to cause congestion in the network to best find the way to consume network resources.
• Alternatively, we might want to achieve the effect of using our network resources efficiently, but avoid congestion altogether.
• Such a strategy is called a congestion-avoidance strategy.

DECBit

• This was an early mechanism to do collision avoidance, developed by DEC.
• In this scheme each router monitors the load it is experiencing and explicitly notifies the end nodes when congestion is about to occur.
• This is done by setting a binary congestion bit in the packets that flow through the router. Hence the name DECBit.
• The destination host copies this bit into the ACK and when this ack is seen by the sender it adjusts its window accordingly.
• More specifically, a router sets its DECbit if its average queue length is greater than or equal to 1.
• A sending increases its queue size by 1 if few than 50% of the acks it received for the last windows worth of data had the DEcbit on.
• Otherwise, it decreases its window size by 0.875.

Random Early Detection (RED)

• In this scheme a router maintain a running average of its queue lengths.
• It also maintains two variables MinThreshold and MaxThreshold.
• If a packet arrives and the average queue length is less than MinThreshold it just queues the packet.
• If its between MinThreshold and MaxThreshold it drops the packet with probability p
• Otherwise, it drops the packet.
• The second case, would cause the sender to timeout earlier than if RED was not used, and so if TCP was being employed the CongestionWindow would stabilize to a workable value without putting as much load on the network.
• When RED is used, the probability p is typically a function of where between MinThreshold and MaxThreshold the average length currently is.