Performance Models and Systems Optimization for Disk-Bound Applications

Purdue University, West Lafayette IN

Investigators

Abstract

Disk-array performance is the primary bottleneck for data-intensive applications, and significant challenges remain in system modeling, algorithm design, and performance optimization of data-intensive applications. Existing analytical models do not incorporate application characteristics, internal disk behavior, and I/O interconnection network contention. The PIs propose to develop an integrated execution-driven simulation environment that incorporates optimization approach driven by application characterization. Unlike traditional worst-case studies that assume adversarial request sequences and lookahead, the proposed approach captures the predictability of the request sequences of real applications, and is expected to lead to the development of better algorithms for applications with large data sets that depend on secondary storage. The proposed environment will be based on a unified and flexible disk-array access model that provides a simple framework for algorithm designers to reason about the complexity of algorithms on disk arrays by incorporating contention on the interconnection network between disks and memory, and consider the complex interactions within a disk access to enhance accuracy over traditional models that assume a flat unit cost per disk block access.