Earticle

As the web game market grows, ensuring service stability through load testing has become increasingly important. Web games comprise a variety of functions with distinct internal logic, ranging from simple data retrieval to complex transaction processing. Therefore, a comparative performance analysis of load-testing tools that accounts for these functional characteristics is crucial for achieving reliable and efficient service operation. This paper evaluates the performance of four widely used load-testing tools—JMeter, k6, Gatling, and Locust—under representative web-game workloads. To emulate realistic database read and write patterns, we implement the core server logic of the web game Pokerogue in a cloud environment rather than simply issuing HTTP requests. We classified workload patterns into write-intensive, read-intensive, and mixed types using distributed tracing, and measured request-generation capability and system resource consumption across five key game functions. Our experiments show that each tool demonstrates distinct strengths depending on the characteristics of individual web-game functions. Specifically, k6 demonstrated high request-generation performance in write-intensive scenarios, while JMeter showed strong performance in read-oriented tasks; Gatling exhibited efficient memory usage in mixed workloads, whereas Locust proved suitable for resource-constrained environments. These results indicate that the selection of a load-testing tool should be informed not only by its request-generation performance but also by the workload characteristics of the target game function. By systematically analyzing function-specific workload patterns together with the performance and resource-usage behavior of each tool, this study aims to provide empirical evidence that can be usefully applied in practical load-testing workflows for web-game services.

Understanding the characteristics of cloud workloads is the key to making optimal configuration decisions and improving the system throughput. However workload characterization of cloud, especially in a large-scale production environment, has not been well studied yet. To gain insights on cloud workloads, we collected a one-week workload trace from a 100-node cloud cluster which hosts 1082 virtual machines. We characterized the workload at the granularity of virtual machines and physical nodes, respectively. We concluded with a set of meaningful observations. The results of workload characterization are representative and generally consistent with cloud cluster for public IaaS service providers, which can help other researchers and engineers understand the performance and VM characteristics of the cloud in their production environments.

Web search distributes query processing across many servers, and each web search server processes multiple queries concurrently to achieve high throughput. A fundamental requirement of web search is responsiveness, which is typically guided by a response time SLA. As this is applied to each and every search query, understanding how system resources are utilized during query execution is important. In this paper, we present workload characterization for web search, with a focus on temporal and spatial locality of index data access. A key finding includes high temporal locality, suggesting that data caching plays an important role in achieving good performance. Moreover, this paper analyzes how effective the preferching of index data is in web search. The prefetching is a very common optimization in web search in order to amortize delay related to data reads from external storage. From the analysis, we observe that the prefetching is not only frequent, but also aggressive, with large size of data to be issued for prefetching. Surprisingly, this optimization is effective in web search. However, some portion of the prefetched data is not accessed by the query due to early termination, and this paper shows that this accounts for 8.3% of total prefetched data. As a future work, we will study how to reconcile many prefetching requests issued by concurrent queries under heavy loads. Lastly, we believe that this work will fuel future endeavors on improving memory management for web search workloads.