1. TCP behavior of a Busy Internet Server: Analysis and Improvements Y2K Oct.10 Joo Young Hwang Computer Engineering Research Laboratory KAIST. EECS.

2. Contents Introduction Data Collection and Web server Single connection behavior analysis Loss Recovery and Receiver window Ack compression Parallel connection behavior Throughput and loss distribution Congestion control analysis Right-edge protocol:Improving single connection performance TCP-INT: Integrated Congestion Control/Loss Recovery Comparison with Application-level multiplexing TCP-INT simulation results Conclusions Critiques

3. Introduction Current HTTP protocol characteristics short TCP connection multiple simultaneous TCP connections Single connection behavior Loss recovery Receiver bottleneck Ack compression’s effect Parallel connection behavior Throughput – relation with # connections Congestion control – relation between packet loss - # connections Loss behavior – distribution of losses across multiple connections

4. Data Collection and Web server IBM RS/6000, 16Mbit/sec token ring network Connected to internet via 4 US Network Access Points(NAPs) TCP emulation engine Load balancing router Trace collection machine Retransmission report

5. Loss Recovery and Receiver window TCP loss recovery is not effective! Sack TCP is not sufficient. New proposed recovery

6. Ack compression ack routerreceiversender

7. Throughput and loss distribution

8. Congestion control analysis Decrease rate : less than single connection Increase rate : n times faster than single connection. Half the total # connections see loss

9. Right-edge protocol: Improving single connection performance No dup. Acks. received One or more dup acks. Received. – insufficient dup acks. Total coarse timeouts Right-edge recovery -send new segment when receive each dup ack.

10. TCP-INT: Integrated Congestion Control/Loss Recovery Dividing TCP functionality to two parts. Part 1 : Reliable, ordered byte-stream abstraction of TCP Part 2 : Congestion control and data-driven loss recovery This part is done in an integrated manner across the set of parallel connections, called TCP-Int. TCP-Int. Single congestion window for the set of TCP connections The combined congestion window is halved on a packet loss New connections do not have to undergo slow-start Integrated data-driven loss recovery

11. Comparison with Application- level multiplexing Application-level multiplexing P-HTTP (Persistent HTTP) SCP (Session Control Protocol) MUX Drawbacks Rewriting existing applications both at the server and the client. Do not allow multiplexing of transfers initiated by more than one application. Undesirable coupling between data transfers that are logically independent. No standard

12. TCP-INT simulation results timeout Slower increase rate. No Timeouts and sharing equal BW

13. Conclusions TCP Packet analysis on a busy HTTP server Single connection behavior Almost 50% of all losses required a coarse timeout to recover. Fast retransmissions recovered from less than 45% of all losses. The socket buffer size limited the throughput of approximately 14% of all observed connections. Ack compression is correlated with subsequent packet losses. Parallel connection behavior A client using parallel connections is a more aggressive user than an application that uses a single connection. Connections with small outstanding windows could experience a larger number of losses. Proposed Right-edge recovery TCP-INT

14. Critiques Incomplete discussion of ack compression Do not propose any way of using ack compression analysis