Workload characterization and business-oriented performance improvement techniques for online auction sites

The workload of auction sites has unique characteristics that do not exist in other e-commerce sites. These characteristics include significant peaks in demand and workload intensity at most critical times such as auction closing times. Workload changes based on time of day along with intense bidding activity towards the closing times of auctions can create surges in the load seen by auction sites as an auction's closing time approaches. If bidders are not able to successfully bid at the very last moments of an auction because of site overload, the final price may not be as high as it could be and sellers, and consequently the auction site, may lose revenue. This dissertation demonstrates that it is possible to design resource management techniques that significantly improve the performance and revenue throughput (i.e., dollars/sec received by an auction site) without increasing the IT infrastructure costs.; As part of a comprehensive study to understand the nature of workloads on auction sites, a very detailed workload characterization was performed using real data from a large production online auction site. Some of the findings from this study include (i) A very large percentage of auctions have a relatively low number of bids and bidders and a very small percentage of auctions have a high number of bids and bidders. (ii) There is some bidding activity at the beginning stages of an auction. This activity slows down in the middle and increases considerably after 90% of an auction's life time has elapsed. (iii) Prices rise faster in the first 20% of an auction's life time than in the next 70% of its life time. However, after the age of an auction reaches 90%, prices increase much faster than in the two previous phases. (iv) Relatively few users are responsible for winning the majority of auctions. (v) Relatively few sellers are responsible for creating the majority of the auctions. (vi) The majority of bids are placed by a relatively small number of unique bidders. Driven by the findings from the workload characterization study, three different resource management techniques for auction sites were developed and their effectiveness was tested by simulation and evaluated on a benchmark auction site test bed. The test bed is a modified version of RUBiS (Rice University Bidding System) that has a workload generator that reflects the workload characteristics elicited in this research.; The first technique is based on the fact that too many auctions closing at the same time on the site can lead to load spikes, since bidding activity along with other activity is expected to be intense during closing minutes of an auction. Results from queuing theory indicate higher coefficients of variation of inter-arrival times of requests result in higher average waiting times at servers. A closing time rescheduling algorithm was proposed that uniformly distributes the activity on auction sites, thereby reducing the coefficient of variation of inter-arrival times, and improving the quality of service to the users by providing faster service. At the time of creation of a new auction, this solution recommends an optimal closing time within a 30-minute window around the original requested time. Trace-driven simulation using actual data from an auction site showed that load spikes can be significantly reduced with this technique. In addition, results showed that rejected requests because of server overload can be reduced by 5% to 8% and timeouts at the server can be reduced by 12% to 25%.; The next solution was driven by the popularity analysis in the workload characterization study, which in summary indicated that a small percentage of auctions attract a large number of bids. In addition to the popularity of auctions, bidder and seller popularity also follow power laws, making caching techniques very effective. This dissertation proposed server-side activity-based caching techniques to reduce server side latencies and improve the performance. The ca...