An Experimental Evaluation of Smart Disk Architectures Using DSS Commercial Workloads

Gokhan Memik, Mahmut T. Kandemir and Alok Choudhary

Abstract

Smart disk systems with large storage capacities and growing computational power are becoming increasingly attractive. The idea is to perform parallel and filtering-type of data intensive computations on disks, close to data, thereby offloading the host processor and increasing the aggregate system power. In this paper, we evaluate in detail the performance of a host-based architecture and a smart disk system using queries from Decision Support System (DSS) databases. We show how to optimize the execution of the whole query being implemented by bundling frequently occurring patterns together and posting the bundle to the smart disks in a single invocation that reduces unnecessary communication between host and smart disks. Using an accurate simulator that can simulate the host-based as well as smart-disk system, we illustrate that the smart disk architecture brings substantial benefits in terms of overall query execution times. In particular, in our base configuration, we are able to achieve improvements over the traditional architecture by as much as 83.5% using representative queries from the TPC-D benchmark suite.

Gzipped Postscript version of the paper