Ilpo Järvinen defends his PhD thesis on Congestion Control and Active Queue Management During Flow Startup

On Thursday the 14th of November 2019, M.Sc. Ilpo Järvinen will defend his doctoral thesis on Congestion Control and Active Queue Management During Flow Startup. The thesis is a part of research done in the Department of Computer Science and in the Wireless Internet research group at the University of Helsinki.

M.Sc. Ilpo Järvinen defends his doctoral thesis Congestion Control and Active Queue Management During Flow Startup on Thursday the 14th of November 2019 at 12 o'clock noon in the University of Helsinki Chemicum building, Room A129 (A.I. Virtasen aukio, 1st floor). His opponent is Professor Anna Brunström (Karlstad University, Sweden) and custos Professor Sasu Tarkoma (University of Helsinki). The defence will be held in English.

The thesis of Ilpo Järvinen is a part of research done in the Department of Computer Science and in the Wireless Internet research group at the University of Helsinki. His supervisors have been Lecturer Markku Kojo and Professor Sasu Tarkoma (University of Helsinki).

Congestion Control and Active Queue Management During Flow Startup

Transmission Control Protocol (TCP) has served as the workhorse to transmit Internet traffic for several decades already. Its built-in congestion control mechanism has proved reliable to ensure the stability of the Internet, and congestion control algorithms borrowed from TCP are being applied largely also by other transport protocols. TCP congestion control has two main phases for increasing sending rate. Slow Start is responsible for starting up a flow by seeking the sending rate the flow should use. Congestion Avoidance then takes over to manage the sending rate for flows that last long enough. In addition, the flow is booted up by sending the Initial Window of packets prior to Slow Start.

There is a large difference in the magnitude of sending rate increase during Slow Start and Congestion Avoidance. Slow Start increases the sending rate exponentially, whereas with Congestion Avoidance the increase is linear. If congestion is detected, a standard TCP sender reduces the sending rate heavily. It is well known that most of the Internet flows are short. It implies that flow startup is a rather frequent phenomenon.  Also, many traffic types exhibit an ON-OFF pattern with senders remaining idle for varying periods of time. As the flow startup under Slow Start causes exponential sending rate increase, the link load is often subject to exponential load transients that escalate in a few round trips into overload, if not controlled properly. It is true especially near the network edge where traffic aggregation is limited to a few users.

Traditionally much of the congestion control research has focused on behavior during Congestion Avoidance and uses large aggregates during testing. To control router load, Active Queue Management (AQM) is recommended. The state-of-the-art AQM algorithms, however, are designed with little attention to Slow Start. This thesis focuses on congestion control and AQM during the flow startup. We explore what effect the Initial Window has to competing latency-sensitive traffic during a flow startup consisting of multiple parallel flows typical to Web traffic and investigate the impact of increasing Initial Window from three to ten TCP segments. We also highlight what the shortcomings are in the state-of-the-art AQM algorithms and formulate the challenges AQM algorithms must address to properly handle flow startup and exponential load transients. These challenges include the horizon problem, RTT (round-trip time) uncertainty and rapidly changing load. None of the existing AQM algorithms are prepared to handle these challenges. Therefore we explore whether an existing AQM algorithm called Random Early Detection (RED) can be altered to control exponential load transients effectively and propose necessary changes to RED. We also propose an entirely new AQM algorithm called Predict. It is the first AQM algorithm designed primarily for handling exponential load transients.

Our evaluation shows that because of shortcomings in handling exponential load transients, the state-of-the-art AQM algorithms often respond too slowly or too fast depending on the actual RTT of the traffic. In contrast, the Predict AQM algorithm performs timely congestion indication without compromising throughput or latency unnecessarily, yielding low latency over a large range of RTTs. In addition, the load estimation in Predict is designed to be fully compatible with pacing and the timely congestion indication allows relaxing the large sending rate reduction on congestion detection.

Availability of the dissertation

An electronic version of the doctoral dissertation is available on the e-thesis site of the University of Helsinki at

Printed copies will be available on request from Ilpo Järvinen: