Carnegie Mellon University Parallel Data Lab Technical Report CMU-PDL-05-106, August, 2005.
Michael Mesnier*, Matthew Wachs, Gregory Ganger
Parallel Data Laboratory
Electrical and Computer Engineering
Carnegie Mellon University
Pittsburgh, PA 15213
*Intel Corp. and Carnegie Mellon
Relative fitness modeling is a new approach for predicting the performance and resource utilization of a workload when running on a particular storage device. In contrast with conventional device models, which expect device independent workload characteristics as input, a relative fitness model makes predictions based on characteristics measured on a specific other device. As such, relative fitness models explicitly account for the workload changes that almost always result from moving a workload across storage devices—for example, higher I/O performance usually leads to faster application execution which results in higher I/O rates. Further, relative fitness models allow service observations (e.g., performance and resource utilizations) from the measured device to be used in making predictions on the modeled device—such observations often provide more predictability than basic workload characteristics. Overall, we find that relative fitness models reduce prediction error by over 60% on average when compared to conventional modeling techniques.
KEYWORDS: storage device, self-managing, CART models, regression trees
FULL PAPER: pdf