1. End to End Internet Packet Dynamics Vern Paxson University of California, Berkeley Presented by Kiran Komaravolu

2. End to End characteristics of Internet paths. Network Pathologies Network Pathologies Out of order delivery Out of order delivery Packet replication Packet replication Packet corruption Packet corruption Bottleneck Bandwidth Bottleneck Bandwidth Packet Loss Packet Loss Loss rates Loss rates Data vs. Ack loss Data vs. Ack loss Loss Bursts Loss Bursts TCP retransmission TCP retransmission Packet Delays Packet Delays

3. Out of order delivery is fairly prevalent Out of order delivery is fairly prevalent Re-ordering is asymmetric Re-ordering is asymmetric Route-Fluttering Route-Fluttering Site-dependant Site-dependant Impact of re-ordering is not very significant. Impact of re-ordering is not very significant. Out of Order Delivery

4. Packet Replication Packet Replication Not very common. Not very common. Link level retransmissions are one main reason. Link level retransmissions are one main reason. Packet Corruption. Packet Corruption. Evidence suggests data packet corruption rate of 0.02%. Evidence suggests data packet corruption rate of 0.02%. Pure acks less prone to packet corruption. Pure acks less prone to packet corruption. TCP provides a 16-bit checksum. Thus it can lose 1 in 65,536 packets to errors. TCP provides a 16-bit checksum. Thus it can lose 1 in 65,536 packets to errors. On average 65,536 x 5000 = 300 million packets will cause an bad packet to be accepted. Paxson argues this is too high. On average 65,536 x 5000 = 300 million packets will cause an bad packet to be accepted. Paxson argues this is too high.

5. Bottleneck bandwidth Self-interference time constant Self-interference time constant Qb = packet size / bottleneck bandwidth. Qb = packet size / bottleneck bandwidth. Packet pair Algorithm. Packet pair Algorithm. Two packets are sent with interval Δts < Qb. Two packets are sent with interval Δts < Qb. When they arrive at receiver Δtr = Qb. When they arrive at receiver Δtr = Qb. ICMP echo packets could be used to measure bandwidth with this idea. ICMP echo packets could be used to measure bandwidth with this idea. Difficulties Difficulties Out of order delivery Out of order delivery Clock resolution limitation Clock resolution limitation Change in Bottleneck bandwidth Change in Bottleneck bandwidth Multichannel links Multichannel links

6. Packet Loss Loss rates doubled in 1995 (from 2.7% to 5.2 %) Loss rates doubled in 1995 (from 2.7% to 5.2 %) Bigger windows do not make any impact on Loss rates. Bigger windows do not make any impact on Loss rates. Loaded and Unloaded pkts Loaded and Unloaded pkts Loaded pkts are those which had to wait for a pending transmission to complete. Loaded pkts are those which had to wait for a pending transmission to complete. About 2/3 rd of all packets are loaded. About 2/3 rd of all packets are loaded. Loaded packets are more prone to be lost. Loaded packets are more prone to be lost. Acks are more prone to be lost than unloaded pkts. Acks are more prone to be lost than unloaded pkts. Packet loss maybe asymmetric. Packet loss maybe asymmetric.

7. Packet loss Losses occur in bursts. Losses occur in bursts. TCP retransmissions TCP retransmissions Transmitting TCP may retransmit unnecessarily Transmitting TCP may retransmit unnecessarily Because all acks were lost. Because all acks were lost. Coarse feedback, sequence “holes” not known to sender. Coarse feedback, sequence “holes” not known to sender. Timed out too early. Timed out too early. Significant portion of retransmissions are “redundant”. Significant portion of retransmissions are “redundant”.

8. Packet Delays Timing Compression Timing Compression Timing Compression occurs when a flight (burst) of packets sent over a period Ts arrives at the receiver over an interval Tr and Tr < Ts. Timing Compression occurs when a flight (burst) of packets sent over a period Ts arrives at the receiver over an interval Tr and Tr < Ts. Zhang et al predicted that ack compression could occur is a flight arrives at a router and faces no cross traffic. (router is under-utilised) Zhang et al predicted that ack compression could occur is a flight arrives at a router and faces no cross traffic. (router is under-utilised) E = (Tr + Cr) / (Ts – Cs) E = (Tr + Cr) / (Ts – Cs) Group is compressed if E < 0.75. Group is compressed if E < 0.75. Most connections experience at least one compression event. Most connections experience at least one compression event. Ack compression occurs more often for dup-acks. (they are sent with less spacing between them.) Ack compression occurs more often for dup-acks. (they are sent with less spacing between them.)

9. Packet Delays Compressed acks advance sender TCP window at a faster rate. May lead to network stress. Compressed acks advance sender TCP window at a faster rate. May lead to network stress. Bandwidth measurement techniques would fail with ack compression. Bandwidth measurement techniques would fail with ack compression. Data Packet Timing Compression Data Packet Timing Compression Can occur due to sudden advance in receiver advertised window. Can occur due to sudden advance in receiver advertised window. Bottleneck will spread out the packets. Bottleneck will spread out the packets. Packets “may” still be compressed after the bottleneck. Packets “may” still be compressed after the bottleneck. Data pkt timing compression is rarer than ack compression. Data pkt timing compression is rarer than ack compression.

10. Available Bandwidth B = (N/W resources used by the connection/ (resources it used) + (resources competitors use). B = (N/W resources used by the connection/ (resources it used) + (resources competitors use). B is a figurative expression for available bandwidth. B = 1 means entire bottleneck was available. B is a figurative expression for available bandwidth. B = 1 means entire bottleneck was available. Internet connections have been found to encounter a broad range of available bandwidth. Internet connections have been found to encounter a broad range of available bandwidth.

11. Q’s ?