Self-adaptive utility-based web session management

Abstract

In the Internet, where millions of users are a click away from your site, being able to dynamically classify the workload in real time, and predict its short term behavior, is crucial for proper self-management and business efficiency. As workloads vary significantly according to current time of day, season, promotions and linking, it becomes impractical for some ecommerce sites to keep over-dimensioned infrastructures to accommodate the whole load. When server resources are exceeded, session-based admission control systems allow maintaining a high throughput in terms of properly finished sessions and QoS for a limited number of sessions; however, by denying access to excess users, the website looses potential customers.

In the present study we describe the architecture of AUGURES, a system that learns to predict Web user’s intentions for visiting the site as well its resource usage. Predictions are made from information known at the time of their first request and later from navigational clicks. For this purpose we use machine learning techniques and Markov-chain models. The system uses these predictions to automatically shape QoS for the most profitable sessions, predict short-term resource needs, and dynamically provision servers according to the expected revenue and the cost to serve it. We test the AUGURES prototype on access logs from a high-traffic, online travel agency, obtaining promising results.

Introduction

According to the Internet World Stats [19], there are over 1 billion Internet users, and when online, they are a click away from any site. Certain events such as news, world events, or promotions announced in mass media or in the Web can cause a flock of users to visit related websites, creating peak loads in minutes. Furthermore, current Web applications rely on technologies such as XML-based Web services for B2B communication, SSL for security, media streaming, and AJAX for interactivity. While these technologies enhance users’ experience and privacy, they also increase the demand for CPU and resources on the server side. As Web applications become more resource-intensive and the number of potential visitors increases, system overload incidence is also growing along.

Scaling the infrastructure of a website might not be simple; for cost reasons, scalability problems, or because some peaks are infrequent, websites may not be able to adapt rapidly in hardware to user fluctuations. When a server is overloaded, it will typically serve no connection, as connections compete for resources. System administrators might get warnings from resource-monitoring services, but in general they get to the problem when the situation has already occurred, and controlling the arrival of users is out of their reach. To address this issue, session-based admission control systems [3], [6] are used to keep a high throughput in terms of properly finished sessions and QoS for a limited number of sessions. However, by denying access to excess users, the website loses potential revenue from customers.

In this study we propose a new approach to this problem which uses learning techniques applied to past data, to create a model for anonymous Web user behavior in a real, complex website that does experience the overload situations mentioned above. The model is then used to support decisions regarding the allocation of the available resources, based on a utility-related metrics. The learning phase captures the selected features in a model according to a utility goal. As a proof of concept, we have selected the features that make a customer more likely to make a purchase, and therefore more attractive – from the point of view of maximizing revenues – to keep in the system in the case of a severe overload. In most of the article, we take as a metric the number of completed sessions that end in purchase; Section 6 describes other possible metrics and applications. In this sense, we are proposing a per-user adaptive utility-based session management.

We describe the architecture of the prototype we have developed, named AUGURES, and the simulation experiments we have performed on weblogs from the servers of a high-traffic online travel agency, Atrapalo.com. The experiments indicate that using AUGURES to prioritize customer sessions can lead to increased revenue in at least two situations: one, when overload situations occur; that is, the incoming transaction load exceeds the site’s capacity and some sessions will have to be queued, redirected to a static site, or dropped; for this study, these should be mostly non-buying sessions, while we try to admit most buying ones. The second scenario is that in which keeping a server running has a quantifiable cost; in this case, one could try to group buying sessions on a small number of servers, possibly shutting down those other servers that would produce little or no revenue.

Our preliminary experiments [13] showed that the users’ intentions for visiting a site can be predicted to some extent from their navigation clicks, results of previous visits, and other session information. In this context, a revenue-based admission control policy can help to avoid revenue loss due to randomly delaying or dropping excess connections. Defining admission policies based on information generated from user behavior models can contribute to devising cost-effective infrastructures, and seems to have promising applications for resource management for medium to large Web infrastructures.

In this paper we present a method for learning, from the analysis of session logs, how to assign priorities to customers according to some metric – in this study, to their purchasing probability in the current session. Our approach consists in using the Web server log files to develop learning models that make predictions about each class of user future behavior, with the objective of assigning a priority value to every customer based on the expected revenue that s/he will generate, which in our case essentially depends on whether s/he will make a purchase. Our learning methods combines static information (time of access, URL, session ID, referer, among others) and dynamic information (the Web graph of the path followed by the user), in order to make predictions for each incoming Web request.