1. CATNIP – Context Aware Transport/Network Internet Protocol Carey Williamson Qian Wu Department of Computer Science University of Calgary

2. Why CATNIP Layered protocol stacks Good: providing a unifying framework Bad: compromise performance vs. Physical Link Network Transport Application

3. Why CATNIP (Cont’d) Observations in Web data transfer using TCP/IP  Poor protocol interactions;  TCP’s window-based flow control mechanism produces data bursts;  Not all packet losses are created equal. Packet losses are costly for small document transfer; Not all packet losses are created equal. Packet losses are costly for small document transfer;  A TCP source has limited control over packet loss effects;  An IP router has significant control over packet loss effects.

5. Design of CATNIP Can we make the TCP/IP protocols “smarter” about the specific job?  Convey application-layer context information to the TCP and IP layers Network Transport Application Document Size Packet Priority

6. Design of CATNIP (Cont’d) Adding context-awareness to TCP:  Rate-Based Pacing of the Last Window (RBPLW)  Early Congestion Avoidance (ECA)  Selective Packet Marking (SPM): Use the reserved high-order bit in the TCP header to convey packet priority information

8. Design of CATNIP (Cont’d) Adding context-awareness to IP:  CATNIP-Good  CATNIP-Bad  CATNIP-RED: RED + CATNIP-Good

9. Evaluation of CATNIP Evaluation Simulation: ns-2 Emulation: use WAN emulation to test a prototype implementation of CATNIP in the Linux kernel of an Apache Web server.

10. Evaluation using simulation Network model: Client 100 Server 1 Server 2 Server 10 Client 1 Client 2 Client 99 1.5 Mbps, 5 ms 10 Mbps, 5 ms RouterSRouterC

11. Evaluation using simulation (Cont’d) Web workload model:  10 Web pages  Use empirically-observed distribution to determine the size, and the number of embedded images

12. Evaluation using simulation (Cont’d) Factors and Levels: Performance metrics:  the transfer time for each Web page  the average packet loss

13. Simulation results DropTail routers:  Mean and standard deviation of transfer times Reno/ RBPLW Reno ECA ECA/RBPLW

14.  Packet loss:  Observations: TCP endpoint control algorithms have little advantage to offer.

15. Simulation results (Cont’d) CATNIP-Good routers:  Mean and standard deviation of transfer times Reno/SPM/Good Reno/DropTail Reno/SPM/RBPLW/ Good ECA/SPM/GoodECA/SPM/RBPLW/ Good

16.  Packet loss:  Observations: Adding context-awareness at the IP routers improves the mean Web page transfer times and the standard deviation of the transfer times. The average packet loss rates with CATNIP-Good are higher than for the DropTail routers.

17. Simulation results (Cont’d) CATNIP-Bad routers:  Mean and standard deviation of transfer times Reno/DropTail Reno/SPM/Bad ECA/SPM/Bad

18.  Packet loss:  Observations: Packet losses are shifted to the high priority TCP packets, that is, throw away the “wrong packet” at the “wrong time”, therefor makes matters worse.

19. Simulation results (Cont’d) CATNIP-RED routers:  Mean and standard deviation of transfer times Reno/DropTail Reno/RED Reno/SPM/CATNIP- RED ECA/RED ECA/SPM/CATNIP- RED

20.  Observations: Reno and ECA perform similarly in almost all cases. The effect of CATNIP-RED is greater than the effect of ECA.

21. Experimental Implementation and Evaluation Experimental environment:  WAN emulator: IP-TNE (Internet Protocol Traffic and Network Emulator)  Web server: Apache Web server (version 1.3.19-5) runs on top of modified Linux 2.4.16 kernel.  Implementation focused on the SPM feature only

22. Client 100 Primary Factor: buffer size of the bottleneck link (64 KB -- 512 KB) 10 Mbps, 5 ms Endpoint Client 1 Client 2 Client 99 1.5 Mbps, 5 ms 10 Mbps, 5 ms Network model Server WAN Emulation RouterSRouterC

23. Evaluation results:

24. Conclusions Not all packet losses are created equal; A TCP source alone has limited control over Web data transfer performance, even with application-layer information; The IP layer has a significant influence on Web data transfer performance, particularly when application- layer context information is available; A simple change to the TCP/IP stack implementation can provide the context information; Changes to the queue management at routers can provide significant performance advantages for the context-aware TCP/IP.