MIT Department of Electrical Engineering & Computer Science

DATE: THURSDAY, MARCH 9, 1995

TIME: TALK AT 2:15
Refreshments at 2:00

PLACE NE43-518

Using System-Level Models to Evaluate Storage Subsystem Performance

Gregory R. Ganger
University of Michigan

Abstract

This talk introduces the use of system-level simulation for evaluating the performance of storage subsystem designs. The standard design-stage methodologies, simulation and/or analytic modeling of the storage subsystem, are fundamentally deficient. Even the most detailed and accurate storage subsystem models suffer from two major short-comings: (1) There is no way (in general) to translate changes in storage subsystem performance into changes in overall system performance. (2) Performance predictions generally do not accurately reflect feedback effects between individual request completions and subsequent request arrivals.

Both problems exist because storage subsystem models tend to treat all requests as equally important. In reality, different requests affect system performance (and subsequent request arrivals) in different ways and to different degrees.

Instead of evaluating storage subsystem designs in a vacuum (or trying to recreate request interactions indirectly), I propose the use of system-level simulation models in which the conventional storage subsystem model is one (very important) component. Other system components (e.g., CPUs, main memory, interrupt controllers, operating system functions) are also modeled at appropriate levels of detail.

In the talk, I will describe various implementation trade-offs related to system-level simulation models. Using a prototype system-level simulator, I will investigate several important storage subsystem design issues, including disk request scheduling and write-back disk cache destaging. These case studies represent instances where the conventional methodologies lead to incorrect results, indicating the importance of system-level modeling.

HOST: Prof. Barbara Liskov