1. 1 School of Computing Science Simon Fraser University CMPT 771/471: Internet Architecture & Protocols TCP-Friendly Transport Protocols

2. 2 Motivations  Congestion Control  Prevents congestive network collapse  Improve Quality of Service (QoS) for UDP  Control transmission rate  Fair bandwidth utilization  Prevents starvation of TCP traffic  ~95% of Internet traffic are TCP  In short, a TCP-Friendly protocol based on UDP

3. 3 Approaches  Resources Reservation  Impossible to know exact bandwidth  May leads to over-allocate  Priority Mechanisms  Require supports by path routers  Adaptive Sending Rate  Easy to implement, application level  Can adapt to changes in bandwidth availability  Improve QoS through loss reduction

4. 4 Adaptive Sending Rate  Congestion Control  Achieved by varying sending rate  12+ algorithms for calculating transmission rate  Main idea is to use TCP throughput model  All claims to be the most effective  Loss-Delay Based Adjustment Algorithm (LDA)  TCP like approach  Increase sending rate during network under-load  Uses feedback to accurately measure RTT  Intended for video and audio streaming  Network and Operating System Support for Digital Audio and Video (NOSSDAV ‘98)

5. 5 Loss-Delay Based Adjustment Algorithm  Loss-Delay Based Adjustment Algorithm (LDA)  Start with a small value, 10 kb/s  Additive increase rate (AIR)  If packet drop, back off to initial value (10 kb/s)  Increase until similar rate as TCP  Relies on feedback  Uses Real Time Protocol (RTP) on top of UDP  Feedback contains losses and round-trip time (RTT)  How much to increase?  Bf = bandwidth factor  AIR = initial value (10 kb/s)

6. 6 Loss-Delay Based Adjustment Algorithm  How do we calculate ?  r = current transmission rate  b = bottleneck bandwidth  Calculating bottleneck bandwidth  b = probe packet size / gap between 2 probe packets  Two sequential packets with small gap means less delay

7. Understanding TCP  Detecting Congestion  Recall TCP average throughput  Fairness mean we must not exceed  Recall TCP average throughput simplistic model  TCP throughput is inversely proportional to  RTT and square root of packet loss probability p 7

8. 8 Loss-Delay Based Adjustment Algorithm  Calculating RTT  No ACK in UDP  Use feedback report  Where t= arrival time, = time elapsed since last report, and = last received sender report

9. 9 Performance of LDA  LDA and TCP

10. 10 Performance of LDA  LDA scalability

11. 11 In Theory, It Works  Not In Practice  Additional network overhead (RTP)  Additional application level complexity  Rogue UDP process could starve TCP-Friendly protocols  Performance Driven  UDP is intended to be light weight and fast  TCP-Friendly protocols would have to yield to UDP  Developers are lazy, they want the fastest connection with minimal amount of work

12. 12 Summary  TCP-Friendly Protocols  UDP based protocols with congestion control  Able to sense the network and adjust send rate accordingly  Promote fair bandwidth sharing  Prevents Network Collapse  Work together with TCP to balance bandwidth  Fill The Gap  TCP-Friendly protocols can fill the gap between TCP and UDP

13. 13 References  TCP-Friendly, Advanced Networking: Pittsburgh Supercomputing Center, http://www.psc.edu/networking/projects/tcpfriendly/ http://www.psc.edu/networking/projects/tcpfriendly/  D. Sisalem, H. Schulzrinne, “The Loss-Delay Adjustment Algorithm: A TCP-friendly Adaptation Scheme”, Network and Operating System Support for Digital Audio and Video (NOSSDAV ‘98), Cambridge, UK, July 8-10, 1998.  J. Mahdavi, S. Floyd, TCP-Friendly Unicast Rate-Based Flow Control http://www.psc.edu/networking/papers/tcp_friendly.html http://www.psc.edu/networking/papers/tcp_friendly.html  Lorenzo Vicisano, Luigi Rizzo (Pisa) and Jon Crowcroft, TCP-like Congestion Control for Layered Multicast Data Transfer (INFOCOM ‘98).