Modern processors typically have only 4-8 performance counters which can be programmed to measure up to thousands of cycle-level performance events. These events can easily generate large amount of data, which is called central processing unit (CPU) big performance data. However, how to extract value from the big performance data faces many challenges. This paper presents a performance data analysis approach, which builds a performance model by iteratively using the gradient boosting regression tree algorithm and quantifies the importance of the performance events of workloads in cloud to guide their performance optimization.