Performance analysis of parallel composite service-based applications in clouds

When processing composite service application jobs containing parallel tasks, service providers can optimize their quality of services (QoS) based on refined parallelism settings and resource allocation schemes by leveraging analytical models. However, building such analytical models is particularly challenging due to the fact that an accurate model is required to capture the dependence among sequential or concurrent services and predict response time of each service with varying degrees of parallelism (DoPs). Faced with these challenges, we propose a novel multiclass multi-pool analytical model for composite service applications deployed in clouds. Specifically, we consider embarrassingly parallel services, which do not require additional overhead to split tasks into multiple independent subtasks. We first establish a multi-pool queue network that takes into account the dependence among services and analyze task parameters of each service. To optimize the QoS of embarrassingly parallel services, we present a differentiated parallel processing mechanism which can set varying DoPs for tasks. We also propose an original modulating partition method to predict important performance indicators of each service. By leveraging the proposed model, service providers can obtain optimal settings for the DoPs, resource allocation, and the number of cloud servers, to achieve specific performance levels. Through extensive experiments based on the rendering service dataset and Alibaba's open cluster traces, we demonstrate that the proposed model can not only provide accurate prediction results but also significantly reduce jobs' response time by at least 20%.