Operating Systems Research Group
Operating Systems Research on Energy, Reliability and Autonomy

SysMiner (System Miner)

Systems are becoming ever so complex. It is not a trivial task to provide the high performance, dependability, scalability and manageability that are demanded by enterprise customers. In addition, new Internet-based applications have led to highly diverse system workloads and access behaviors, which makes customizing a system for different workloads a more challenging and sometimes a formidable task. The increasing complexity also worsens system dependability. Since modern systems consist of many components that can fail, the failure rate for a system is much higher than before. Another result of the increasing system complexity is that the administrative (maintenance) cost has become a significant part of a system's TCO (Total Cost of Ownership). To maintain a system, skilled IT professionals are required to install, configure, operate, backup and tune it.

To address the above problem, it is desirable to build systems that can "look after themselves". Such systems are called autonomic systems. The main idea of autonomic systems is to have a system manage itself with minimum human intervention. An autonomic system includes several elements:

To provide autonomic capabilities, a system first needs to automatically capture and characterize the occurring storage access behavior, based on which it can change its control policies or configurations to adapt to application workloads. However, without proper analyzing tools, storage access behaviors are difficult to characterize, especially for workloads whose behaviors change dynamically from one time period to another. Typical storage access behaviors include temporal locality, spatial locality, access frequencies, regular access patterns, block correlations, distributions and many other complex patterns.

This project investigates a novel technology called system mining that applies data mining techniques to improve system performance, dependability and manageability. We will use storage systems as our experimental systems to investigate and evaluate our ideas.

Publications: Members: