The Effects of Active Queue Management on Web Performance Annual Conference of the Special Interest Group on Data Communication (SIGCOMM 2003) August 25-29, 2003 Karlsruhe, Germany Presentation: Andreas Haeberlen COMP629, February 26, 2004 © 2004 Andreas Haeberlen, Rice University Long Le Jay Aikat Kevin Jeffay F. Donelson Smith Dept. of Computer Science University of North Carolina (Chapel Hill) 1 Who would you believe? David Clark (MIT) 81-89: Internet Chief Protocol Architect Bruce Davie (Cisco) MPLS Kevin Jeffay (UNC) Real time (YARTOS), QoS, Multimedia ? Deborah Estrin (UCLA) Multicast routing, BGMP, ... Sally Floyd (ICIR) F. Donelson Smith (UNC) Pushback, ECN, High-speed TCP, TFRC, ... Andrew, SNA, FAPL Larry L. Peterson (Princeton) Scout, TCP Vegas, x-Kernel... Craig Partridge (UCLA) RTP, Gigabit Networking ? "AQM is useless without ECN" -- this paper © 2004 Andreas Haeberlen, Rice University Scott Shenker (ICSI) CAN, RSVP, ALM, ... "Thou shalt use AQM" -- RFC2309 2 Introduction Host X Router © 2004 Andreas Haeberlen, Rice University Internet uses packet switching Need queues to handle bursts Queues are of finite length Drop packets when queue is full Overload can lead to congestion collapse Solution: Congestion control mechanism AIMD extension for TCP 3 Introduction: RED X min max pdrop Average queue length © 2004 Andreas Haeberlen, Rice University Problems with 'drop tail': High packet loss Full queue latency Lockout Idea: Drop some packets before queue is full to trigger AIMD mechanism Short bursts should still be tolerated! RED: Use average queue length (EWMA) to compute drop probability 4 Introduction: ARED, PI, REM min max pdrop 1.0 0 Average queue length RED: pb max P (avg min th ) max th min th PI: p(kT ) a (q(kT ) qref ) b (q(( k 1)T ) qref ) p(( k 1)T ) REM: p(t ) max( 0, p(t 1) ( (q(t ) qref )) x(t ) c)) © 2004 Andreas Haeberlen, Rice University RED improvements: Gentle mode Adaptive RED (ARED) Many other variants; >50 proposed in 1999! Biggest problem: RED is either unstable or responds too slowly Application of control theory: Proportional integrator (PI) Random exponential marking (REM) 5 Introduction: ECN 4 IP header IHL DSCP Length flags 16-bit ID Fragm. offset Protocol Checksum TTL Source IP Destination IP Data ... TCP header Offset Flags TCP Window ... Sender ECE: Echo CE © 2004 Andreas Haeberlen, Rice University Receiver Idea: Mark packets instead of dropping them Explicit congestion notification (ECN) Special bits in IP and TCP headers Effect: Send window is reduced CE: Congestion Experienced CWR: Congestion Window Reduced 6 Goal of this paper Study the effects of active queue management on the response times experienced by web users Compare ARED, PI, REM with/without ECN © 2004 Andreas Haeberlen, Rice University 7 Methodology FreeBSD machines (ARED, PI, REM) Traffic generator Server Network models interconnection between two ISPs Traffic generators emulate browsing users Offered load varied between 80%-105% of bottleneck © 2004 Andreas Haeberlen, Rice University 8 Cumulative probability Results: 80% load, no ECN Response time (ms) Offered load: 80% of bottleneck link (100Mbps) Result 1: For offered loads up to 80%, AQM does not provide better response time than drop-tail FIFO Result 2: ARED even decreases performance! © 2004 Andreas Haeberlen, Rice University 9 Cumulative probability Results: 90% load, no ECN Response time (ms) Offered load: 90% of bottleneck (98%, 105% similar) Result 1: For 80% of the responses, PI, REM and drop-tail all provide reasonable performance Result 2: In the remaining cases, PI is better © 2004 Andreas Haeberlen, Rice University 10 Cumulative probability Cumulative probability Results: 98% load, with ECN Response time (ms) Response time (ms) Offered load: 98% of bottleneck Result 1: With ECN, both PI and REM significantly improve response time at offered loads >90% Result 2: Response time with ARED is consistently poor © 2004 Andreas Haeberlen, Rice University 11 Summary and Conclusions Major results: Up to 80% load, drop-tail is as good as AQM Without ECN, PI is slightly better than ARED, REM With ECN, both PI and REM improve response times significantly at 90% load and above ARED consistently performs poorly, even with ECN Conclusions: AQM without ECN: Small improvement AQM with ECN: Significant improvement - good enough to operate links at near saturation levels! © 2004 Andreas Haeberlen, Rice University 12 Review Sound: Realistic experiments, good results, well presented Relevant: Shows a way for providers to get better link utilization (ECN-capable routers) Interesting: Results for RED inconsistent with its high reputation in the community © 2004 Andreas Haeberlen, Rice University 13 The End Rebuttal © 2004 Andreas Haeberlen, Rice University 14 80% Uncongested Argument: Queue management in an uncongested network is not very interesting Purpose is to show what the interesting case is © 2004 Andreas Haeberlen, Rice University 15 ECN helps over 90% Argument: PI paper already has that result Paragraph of prose is not a proof Popular belief about ARED shows that intuition is sometimes misleading What is new? - Experimental validation © 2004 Andreas Haeberlen, Rice University 16 "Everybody loves RED" May, M., Bolot, J., Diot, C., and Lyles, B., Reasons not to deploy RED, technical report, June 1999. "The main results we found were, first, that RED with small buffers does not improve significantly the performance of the network... Second, parameter tuning in RED remains an inexact science, but has no big impact on end-to-end performance." M. Christiansen, K. Jeffay, D. Ott, and F.D. Smith, Tuning RED for Web Traffic, ACM SIGCOMM, August 2000. "We conclude that for links carrying only web traffic, RED queue management appears to provide no clear advantage over tail-drop FIFO for end-user response times... There are some limitations of this study that should be considered... Congestion on both paths on a full-duplex link and over multiple router hops should also be considered." © 2004 Andreas Haeberlen, Rice University 17 ARED is bad Argument: Goal is link utilization, stabilize queue delay Stable, but high queue delay is not desirable Argument: Web traffic results were already published in SIGCOMM 2000 Best paper award for stale results?!? -> See conspiracy theory slide © 2004 Andreas Haeberlen, Rice University 18 RED Deployment Argument: Would Cisco deploy it without proof that it works? If IETF can make an error, so can Cisco Huge companies have made huge mistakes: Microsoft and the Internet IBM and microkernels Proof by Credibility © 2004 Andreas Haeberlen, Rice University 19 RED Parameter Setting Argument: Need more research on parameter setting Many papers have been written on RED - lack of research? There are even summary papers on the family of RED techniques! Even RED authors do not have good recommendations © 2004 Andreas Haeberlen, Rice University 20 Evaluating RED Argument: RED was designed for something else Question: Which metrics matter? Delay for interactive applications certainly matters! © 2004 Andreas Haeberlen, Rice University 21 "Experiment is unrealistic" Web traffic only Focus is response time; web surfing is the dominant interactive application Network too small Need controlled evironment Access to backbone routers? Configuration error Careful calibration; built on previous work Much better than in pro-RED papers © 2004 Andreas Haeberlen, Rice University 22 PI is cool Not the main point of the paper - really advocates use of ECN (if you read between the lines) © 2004 Andreas Haeberlen, Rice University 23 "ECN is not yet deployed widely" Not true - newer OSes do have support for it Paper breaks the 'vicious circle of deployment' by providing a good case for ECN © 2004 Andreas Haeberlen, Rice University 24 Conspiracy: Best paper award Did SIGCOMM PC conspire against RED? Is that plausible, given the list of names on the RED RFC? How is best paper decision made? © 2004 Andreas Haeberlen, Rice University 25 The End Backup © 2004 Andreas Haeberlen, Rice University 26 "Experimental error" "Usually, there were no noticeable differences between repetitions; where there were..." Unrealistic to assume that results from an experiment of this size can be reproduced exactly © 2004 Andreas Haeberlen, Rice University 27 "Experimental error" II "We chose two target queue lengths..." - "a queue size ... that would represent a 'best practice' choice" They did try other parameters Many other authors complain that choosing parameters for DT/RED is not an 'exact science' © 2004 Andreas Haeberlen, Rice University 28 "Experimental error" III "This is an artifact of our traffic generation model wherein browsers generate requests less frequently as response time increases" Is this really unrealistic? Applies only to 105% results; main points can be demonstrated at 90%/98% "The exact reasons for the observed differences remains the subject of continued study." © 2004 Andreas Haeberlen, Rice University 29 "Experimental error" IV "The exact reasons for the observed differences remains the subject of continued study." © 2004 Andreas Haeberlen, Rice University 30 "AQM method X performs better" AQM design space is large © 2004 Andreas Haeberlen, Rice University 31 "What about other protocols? TCP/IP is the dominant protocol in the internet today ARED and the other AQM schemes were specifically designed for TCP/IP © 2004 Andreas Haeberlen, Rice University 32