This
paper presents techniques for measurement and analytic modeling
of disk subsystem performance. We measure disk subsystem
performance using benchmark programs and develop analytic
models of measured systems. The models are calibrated and
used as tools for performance prediction. Our goal is to
show problems related to practical modeling of real life
computer systems and limitations of some traditional modeling
techniques. We present models of load-dependent disks, disk
cache, and load-dependent Mean Value Analysis models having
high predictive power. Our results show that real systems
include highly nonlinear behavior caused by disk optimization
algorithms and by caching, and cannot be satisfactorily
modeled using the traditional load-independent Mean Value
Analysis. We performed experiments for a variety of operating
systems including Windows NT. The development of load-dependent
models, on the other hand, is not simple and requires a
substantial calibration effort. We developed a new quantitative
indicator for evaluation of predictive power of our analytic
models.
The results of this research have been obtained in the Experimental
CS Lab developed using the NSF grant ILI DUE-9751724. Without
this grant our experimental work would not be possible.
