Data-Driven Characterization and Modeling of Web Map System Workload

Every month, billions of users access Web Map Systems (WMSs), such as Google Maps, to visualize geospatial data. A large number of users and the huge amount of data demanded by these applications make the design and development of WMSs a challenging task, especially in terms of performance and scalability. In this context, workload generators become crucial tools, as they help system administrators to plan the capacity of WMSs and design provisioning strategies for peak load scenarios. However, little is known about the workload patterns generated by WMS users. In this work, we use data anonymously collected from sessions of a client application of Google Maps to devise a model that describes how users of desktop terminals navigate in a Web map. Based on this model, we implement a workload generator called MUSeGen. We compare the workload patterns generated by MUSeGen against the workload patterns found in real data. Results show that MUSeGen generates synthetic traces whose navigation patterns closely match those found in real data. We also compare MUSeGen against HELP, a workload generator built upon previous findings on empirical knowledge on the usage of WMSs. Results show that the number of issued operations per session in HELP is, on average, four times lower than that in MUSeGen and the number of tiles requested is, on average, twice lower than that in our tool. In addition, navigation patterns in HELP are much simpler than in MUSeGen. These findings support the conclusion that MUSeGen produces more realistic workloads than HELP. To illustrate how such differences affect performance evaluation in practice, we carry out a performance evaluation of a real WMS under workloads generated by HELP and MUSeGen. Our evaluation shows that the system capacity under HELP is three times less than that obtained under MUSeGen, highlighting the value of MUSeGen.