MIT OpenCourseWare http://ocw.mit.edu 6.033 Computer System Engineering Spring 2009 For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.6.033 Lecture 13 Sam Madden Congestion Control Last time, we saw how the end to end layer ensures exactly once delivery,using windows to efficiently make use of the network. Couple of loose ends: - Reordering -- packets may arrive at the receiver out of order recv(p)slot = p.sno - lastif (slot > last + size)drop else new = slot is empty if (new) put p in slot ack p if (slot == last and new) deliver prefix to app slot += size(prefix)Cumulative acks So far, have presented acks as being selective -- for a specific packet In Internet, sometimes cumulative acks are used in figure above, when p1 arrives, would send ack(2), indicating that all packets up to 2have been received. + insensitive to lost acks (if using selective acks, a dropped ack requires a packet retransmit. Here, if theack for P2 was lost, when P0 arrives, receiver knows P2 arrived.) - sender doesn't know what was lost here, sender has sent p3, p4, and p5; p4 and p5 have been receiver, butonly receives an ack for up to p2. can lead to excessive retransmission In the TCP, both cumulative and selective acks are used: common case, use cumulative, so that a lost ack isn't problematicwhen acking a packet for which later packets have arrived,use selective acks, so that sender knows what has arrived Choosing the window size Last time, we send that the window size should be: Window = Rate x RTT Where Rate is the maximum rate the receiver can accept packets at This will keep the receiver busy all the time E.g., suppose receiver can handle 1 packet / ms , and RTT is10 ms; then window size needs to be 10 packets for receiver toalways be busy 2 questions:1) How does the receiver know how fast it can process packets? 2) What if the network can't deliver packets as fast as the receiver can process? 1) It doesn't. In practice it doesn't matter that much -- just set it to be big enough. Forexample, on 100 Mbit ethernet, with 10 ms RTT, window = 100 Mbit/sec * .01 sec = 1 Mbit = 128 KB So what if the network can't handle packets at this rate -- is that a problem? Suppose sender and receiver can both communicate at 3 pkts / sec, but that there is abottleneck router somewhere between the sender and receiver that can onlysend 1 pkts / sec. Suppose RTT is 2 sec ( so receive window is 6 pkts) After 1 sec, there will be 2 packets in this router's queuesAfter 2 seconds, there will be 4 packets Congestion collapse = delivered load constantly exceeds available bandwidth Notice that if we could adapt RTT on sender fast enough, that would help, since we'd sender fewer duplicates. But we could still get into trouble if we have a big window and there is a slowbottleneck. Why does congestion arise? Routers in the middle of the Internet are typically big and fast -- much faster than theend hosts. So the scenario above shouldn't arise, should it? Issue: sharing. Sources aren't aware of each other, and so may all attempt to usesome link at the same time (e.g., flash crowds). This can overwhelm even the beefiest routers.Example: 9/11 -- everyone goes to CNN. Apple -- releases new products. What do do about it: Avoid congestion:Increase resources? Expensive and slow to react; not clear it will help Admission control? Used in telephone network. Some web servers do this, toreduce load. But its hard to do this inside of the network, since routers don't reallyhave a good model of applications or the load they may be presenting. Congestion control -- ask the senders to slow down. How do senders learn about congestion? Options: 1) Router sends feedback (either directly to the sender, or byflagging packets, which receiver then notices and propagatesin its acks.) Works, but can be problematic if network is congested because thesemessages may not get through (or may be very delayed) 2) Assume that packet timeouts indicate congestion. What to do about congestion: - Increase timeouts. - Decrease window size. Need to do this very aggressively -- otherwise it is easy for congestion to collapse to arise. CC is hard to recover from. TCP does both of these: Timeouts: exponential backoff On retransmitted packet, set: set timeout = timeout * c ; c = 2 up to some maximum timeout increases exponentiallyAfter ack's start arriving (losses stop), timeout is reset based on RTT EWMA (as in last class) 2 windows: receive window + "Congestion Window": Window size = min(congestion window, receiver windoow) On retransmitted packet: CW = CW/2 (Min size = 1) On ack'd packet: CW = CW + 1 (up to receive window size) To initialize window quickly, use TCP "slow start": At the start of the connection, double CW up to receive window, unlesslosses occur: Slow start -- while connection is starting only On each ack: CW = CW * 2 (max receive window) Is TCP "fair"? no ; applications aren't required to use TCP in which case they won't obey congestion rules. Applications can also use multiple TCP connections.So why does this work? Apparently because most traffic is
View Full Document
Unlocking...