RAP: End-to-end Rate Based Control for Real Time Streams in the Internet R. Rejaie et al, Infocom 99Slide 2RAP MechanismRAP (cont)Slide 5Slide 6Slide 7TCP friendlinessSlide 9Slide 10Slide 11Slide 12RAP behavior with REDSlide 14Slide 15RAP behavior with RED (cont)ConclusionsRAP: End-to-end Rate Based Control for Real Time Streams in the InternetR. Rejaie et al, Infocom 99•RAP: Rate Adaptation Protocol•Applications: web server or VOD server streaming•Uses UDP and RTP•Mimics TCP’s AIMD behavior•Main design goal: friendliness to TCP•End to end, application level implementation•Layer encoded, stored real time stream•Source adapts rate by adding/removing layers based on ETE feedbackRAP Mechanism•Receiver individually ACKs packets•It delivers packets to playout buffer even if received out of order•Each ACK carries sequence #•Sender estimates round trip time SRTT from ACKs•Sender keeps packet timers and checks for potential timeoutsRAP (cont)Increase/decrease (AIMD) alg:•Rate adjusted each SRTT•No loss: add one more packet in SRTT•Loss detected: reduce # of packets per SRTT by ½ (like in Reno)•“Cluster loss” (ie, many consecutive packets lost): react only to first loss – similar to TCP SACK behaviorRAP (cont)•RED (Random Early Drop) used to limit the burst loss occurrence•RED improves TCP-friendliness: it allows TCP to catch up with RAPSimulation experimentsNS-2 TCP flows: FTPResources are scaled up proportionally to # of users.FG = fine grain adaptation;It adjusts rate continuouslyTCP friendliness•RAP not very friendly to Tahoe!•Tahoe suffers from frequent time outs and slow-start episodes•Other TCP versions fare better•In following experiments, TCP SACK is usedTCP-SACK in useTCP friendliness•Half flows are RAP, half are TCP•Fairness ratio = avg RAP thrpt/avg TCP thrpt•RAP fair to TCP SACK (except for small # of flows and small round trip delays)RED GatewaysMax p = 0.16;short queue hurts TCP.005; large queueFairness = RAP thr/TCP thrRAP behavior with REDIf max p is too small (eg .005):Avg queue size grows large, packets are tail-dropped; system has large queue fluctuations–With small # of flows, the period is large; TCP recovers less easily than RAP–With large # of flows, the period is shorter; TCP flows are hit less than the evenly spaced RAP flows; RAP performs poorly!RAP behavior with RED (cont)•As max p is increased, more packets are random dropped and the queue becomes more stable •Buffer utilization (and throughput) are lower•TCP congestion window becomes small, and RAP takes advantage of it, “stealing’ the avail bandwidth from TCP flowsConclusions•RAP is reasonably TCP friendly in large scale nets (many flows) and large windows•With a few flows, the scheme becomes unfair•RED improves fairness; but parameters must be properly tuned; bed RED param selection (eg, max p = .005) can harm real time
View Full Document