Research On Software Performance Anomaly Diagnosis Mechanism In Cloud Computing

Cloud service software needs to provide continuous online service. But complexity of software code is increasing with the strengthening of modern software functionality. It is almost impossible to develop a completely bug-free software system. When the software runs in a cloud environment, if a performance anomaly is caused by a bug, it will be difficult to diagnose performance anomalies for developers. On the one hand, these performance bugs only manifest under specific conditions, for example, special user input or system configuration. It is difficult to reproduce these performance bugs outside the runtime environment. Thus, it is hard to use existing offline debugging tools to diagnose these performance bugs. On the other hand, performance bugs often produce no error message that can help developers solve these performance problems. Therefore, the research on the diagnosis mechanism of performance anomaly in cloud environment is important.To address the performance anomaly problems of the service software in the cloud, we propose a performance anomaly diagnosis mechanism for the software in the cloud. Firstly, we design a tracing mechanism of system calls to get the information for diagnosis with low overhead. This tracing mechanism can continuously trace system calls generated by the software during its production run. Then we design a modelling approach to model the system call behavior of the software based on the self-organize map. This model can online diagnose the performance anomaly quickly. Finally, we designs a diagnosis approach to identify abnormal execution units with the behavior model of system calls by comparing the differences between different execution units, and use the majority voting to infer the abnormal system calls related to the performance anomaly.Experimental results show that the software performance anomaly diagnosis system can diagnosis the real performance bugs effectively. This system can infer system calls related to the anomaly within average 7 minutes. This system imparts 2.2% average runtime overload to the tested software.