Applying Data Mining Techniques to Computer Systems

Modern computer systems desire several properties, such as high performance, reliability and manageability. To deliver these properties requires a lot of human effort which is costly and error-prone. In order to automate the procedure in development and management, the first step is to analyze and characterize systems by computers. The system data such as source code, development documents, execution traces, access traces, etc. provide a valuable asset for us to target the solution. In the meantime, the huge amount of data, however, renders a tedious and difficult task on managers and developers, and hence the hidden information would be difficult to extract. During this talk, I will present a novel approach to analyze various system data by applying data mining techniques. This approach can effectively obtain useful information hidden in huge amount of system data, and then such information can be exploited for improving system performance, reliability and manageability. Specifically, I have applied different data mining algorithms on different types of system data such as source code and access traces to achieve different goals including automated debugging and system behavior characterization. The results demonstrate that data mining is an effective and promising method to help us solve problems in computer systems.