Workload characterization of a large ecommerce platform.

Abstract

Several works covered the characterization of web servers’ workload. These works resulted in a compilation of patterns called invariants, i.e., recurrent observations seen in multiple servers. Although some of these works focused on ecommerce systems, they analyzed data from small store servers within a short timespan in the late 90s and early 2000s. Thus, this work proposed a workload characterization of a multinational ecommerce company server and its comparison with the previous invariants found in the literature. We found that some patterns, such as the presence of peaks and valleys in the arrival rate distribution over time and its relation with the working hours of the day, continue to be present in modern ecommerce servers. Meanwhile, others have diminished or disappeared, such as the correlation between the arrival rate and the latency. We also conducted analyses not found in the literature, such as the impact of Black Friday on the server workload and the analysis of two new metrics: surge queue length and spillover count. We found a higher arrival rate during Black Friday than on typical days, a skewed distribution for surge queue length, and an association between the spillover count and high queue length values and latency.