Toward Self-Tuning Storage Systems
As data storage systems become larger, more complex, and support increasingly diverse workloads, tuning and optimizing their performance has become a daunting task. One proposed solution is to create intelligent, self-tuning storage systems that are capable of monitoring, and analyzing and adapting to their workloads. In this talk, I will discuss some of the reasons why building high-performance storage systems is difficult, and then describe a new method by which storage systems can learn to make accurate predictions about several key properties of future files (including access patterns, lifespan, and size) by exploiting the strong and workload-specific associations between a file's properties and the name and attributes assigned to it when it is created. I will illustrate this method with examples and results from several contemporary NFS (Network File System) workloads. I will also discuss how these predictions might be used as part of the design of new file storage systems.