A business-oriented load dispatching framework for online auction sites Daniel A. Menasc´ e Department of Computer Science George Mason University Fairfax, VA 22030, USA menasce@cs.gmu.edu Vasudeva Akula School of Information Technology & Eng. George Mason University Fairfax, VA 22030, USA vakula@gmu.edu Abstract Online auction sites have unique workloads and user be- havior characteristics that do not exist in other e-commerce sites. Earlier studies by the authors identified i) significant changes in the workload depending on time of day and ii) the presence of heavy-tailed distributions involving bidder and seller behavior. Other studies indicate that a small frac- tion or power users, among millions of registered users, contribute to the majority of the site’s revenue. Poor qual- ity of service to power users may imply in loss of business and cause significant loss of revenue to an auction site. This problem could be mitigated by dedicating resources in an exclusive manner to this small group of users in order to im- prove the quality of service they receive. However, this ap- proach can lead to i) under utilization of these dedicated resources and ii) overloading of resources allocated to reg- ular users when the load from power users is low. In this pa- per, we propose a scheme whereby resources are primarily dedicated to power users, but regular users can take advan- tage of these dedicated resources when spare capacity is available. This paper provides a business-oriented frame- work for dispatching requests to the various servers of an online auction site. Our approach uses a controller that can dynamically shift the load to different types of servers as the workload changes. Experimental evaluation showed, among other things, that the total number of bids processed using a dynamic controller to allocate resources can be im- proved under heavy load conditions compared to using a load balancing technique that does not differentiate among users. 1. Introduction Online auction sites have very specific workload charac- teristics as shown in a very detailed study that used data from a major online auction site conducted by the au- thors [3]. One of the numerous conclusions of that work, is the presence of heavy-tailed distributions in many aspects of an online auction’s user behavior and bidding activity. For example, a very large percentage of auctions have a rel- atively low number of bids and bidders and a very small but non-negligible percentage of auctions have a high num- ber of bids and bidders. A large percentage of auctions have a low closing price and a very small but non-negligible per- centage of auctions have a large closing price. The findings of that study allowed us to i) cre- ate an experimental testbed for online auction sites that extends RUBiS: Rice University Bidding Sys- tem (see http://rubis.objectweb.org/)—a benchmark de- signed by Rice University—with a workload generator that is more realistic and follows the distributions ob- served in a major online auction site, ii) design resource allocation schemes that exploit the observed character- istics of the workload and are based on bidding activ- ity in order to increase the revenue of the online auc- tion site. As an example, on a previous study, the au- thors designed and evaluated novel cache placement and replacement policies for server-side caches in on- line auctions [14, 15]. In another example, the authors proposed and evaluated a closing time rescheduling algo- rithm aimed at smoothing out the load peaks observed at real online auction sites as the closing time for an auc- tion approaches [16]. There are periods of time during the day in which the ac- tivity on an auction site is more intense than in others. Auc- tion sites require users to authenticate themselves before performing revenue generating transactions such as bidding and creating auctions. Once the user is authenticated, his- torical activity about that user can be used to provide differ- entiated service to that customer. Typically, a small percent- age of users contribute to a large percentage of revenue for an auction site [6]. We use some of these facts to develop user profiles with the information necessary to provide dif- ferentiated services to address some of the performance is- sues during peak periods on auction sites. We present business-oriented resource allocation poli-