=================================================== From gulati@cs.rice.edu Wed Feb 25 14:14:17 2004 =================================================== The Effects of Active Queue Management on Web Performance This paper presents an evaluation of the effects of various active queue management algorithms on the response time distribution of web users. The main AQM schemes studied are Proportional Integrator(PI), Adaptive Random Early Detection(ARED) and Random Exponential Marking(REM). All the schemes are studied with and without the use of Explicit Congestion Notification(ECN) in TCP header. The authors did an extensive experimental setup consisting of a large number of computers arranged in a manner, so that all the traffic goes through two ISP's operated at 100Mbps. These two ISP's used separate dedicated 100Mbps links for both directions. The computers are either browsers or servers. The traffic generated emulates the web browsing pattern of thousands of users and their response time is used as primary evaluation criterion. Authors find that at loads up to 80% of bottleneck link capacity, simple drop-tail FIFO queue management works as good as other schemes. For higher loads of 90% of link capacity PI does slightly better than other schemes without ECN. With ECN both PI and REM shows substantial improvement, with response times quite close to that achieved in unloaded network. Finally they claim that ARED performs poorly both with and without ECN. I think that paper actually shows that no matter how smart you make the routers, TCP at the end hosts is not too good to take advantage of it and ECN is the only way to make it all work. One thing that would be interesting to see is that how these schemes do with FAST TCP. My conjecture would be that substantial improvement should be observed even without ECN as FAST TCP reacts to queuing delays and not packet loss. -Ajay =================================================== From muhammed@ece.rice.edu Wed Feb 25 21:14:49 2004 =================================================== The results of this paper have a high impact. The paper establishes that AQM schemes have significant benefits only at very high link utilizations (90% and above). But almost all links in the internet core are over commissioned and are currently operating at very low link utilizations (5%). This is to absorb bursts in traffic. Therefore AQM does not have a role at the network core since the link utilizations are way below 90%. A short coming of this paper is that there is almost no explanation of why the system shows such behavior. The paper describes the graphs without telling why they are so and why the different AQM schemes compare like that. Such explanations would help understand how other versions of TCP (say FAST-TCP) would react to AQM. Will they be aggressive and not react at all to AQM drops because they interpret drops by the AQM module as random drops? =================================================== From takhoa@rice.edu Wed Feb 25 22:57:04 2004 =================================================== This paper provides a good understanding of the behaviors of different queue management algorithms. However, I wonder how close the setup represent the true behaviors of the Internet. Although the authors used a web-trace analysis model that was based on real Internet traffics, the model has been built using some theoretic models. Thus, it is possible that this model will be more biased towards theoretic AQM algorithms. Another issue is the experimental network setup. Each host with a 100Mbps link simulates several thousand users. In the uncongested case, the authors allow a 1Gbps link at the routers. But would this create congestion at each browers? Web traffics for thousands of users could be more than 100Mbps. In general, the problem I have with this paper is the fact that it tries to analyze the behaviors of AQM for the real Internet entirely through an experimental testbed. The paper does not make any significant attempt to justify how close or far away the testbed is to reality. It should attempt to do a real traffic analysis between two real geographic points. =================================================== From amsaha@cs.rice.edu Thu Feb 26 17:21:18 2004 =================================================== The paper does study of previously proposed adaptive queue management (AQM) techniques (Proportional Integrator (PI), Random Exponential Marking (REM) and Adaptive Random Early Detection (ARED)) . The AQM techniques are studied both with and without Explicit Congestion Notification (ECN) modifications to TCP/IP. The authors find that without ECN none of the AQM methods are better than a simple drop-tail FIFO queue management. With ECN, PI and REM performed very well even under very high load on the network. ARED never performed well. I find paper like these very difficult to read. The paper just presents graphs and gets away with it. The authors suggest that probably one of the main reasons why ARED performs worse than REM and PI is because ARED does 'packet mode' monitoring and not 'byte mode' monitoring. However, couldn't the authors simply try the 'byte mode' monitoring and check if their hypothesis was correct. This paper seems to suggest that AQM algorithms are useless without ECN in TCP/IP but I dont find the authors stating that very strongly. Thanks, Amit +----------------------------------------------------------------------+ | Amit Kumar Saha, amsaha@rice.edu, http://www.cs.rice.edu/~amsaha | | Rice University, 6100 Main St, MS-132, Houston, Texas 77005, USA. | +----------------------------------------------------------------------+ =================================================== From santa@rice.edu Thu Feb 26 23:02:35 2004 =================================================== The Effects of Active Queue Management on Web Performance This paper compares Active Queue Management techniques on web performance. It implements and compares drop-tail, ARED, REM and PI AQM methods. They conclude that ARED performs even worse than drop-tail for HTTP loads, and ECN significantly betters the performance of PI and REM. The paper is very well packaged, in the sense that lots of results - some new, but mostly old - are put together nicely and conclusions drawn. The only significant message to carry away from this is the ECN should be worth deploying in the internet. The other results - no AQM performs better than drop-tail for upto 80% traffic load, PI results in improvement than drop-tail, and, ARED is poorly performing with web traffic - are well-known from before. Their evaluation has significant flaws - their traffic consists of solely web traffic which is not realistic, and their topology is also not realistic. It almost seems the paper was out to prove RED is bad. The real point got buried in the paper.