PDL Abstract

IRONModel: Robust Performance Models in the Wild

SIGMETRICS’08, June 2–6, 2008, Annapolis, Maryland, USA.

Eno Thereska*, Gregory R. Ganger

Dept. Electrical and Computer Engineering
Carnegie Mellon University
Pittsburgh, PA 15213

* Microsoft Research


Traditional performance models are too brittle to be relied on for continuous capacity planning and performance debugging in many computer systems. Simply put, a brittle model is often inaccurate and incorrect. We find two types of reasons why a model’s prediction might diverge from the reality: (1) the underlying system might be misconfigured or buggy or (2) the model’s assumptions might be incorrect. The extra effort of manually finding and fixing the source of these discrepancies, continuously, in both the system and model, is one reason why many system designers and administrators avoid using mathematical models altogether. Instead, they opt for simple, but often inaccurate, “rules-of-thumb”. This paper describes IRONModel, a robust performance modeling architecture. Through studying performance anomalies encountered in an experimental cluster-based storage system, we analyze why and how models and actual system implementations get out-of-sync. Lessons learned from that study are incorporated into IRONModel. IRONModel leverages the redundancy of high-level system specifications described through models and low-level system implementation to localize many types of system-model inconsistencies. IRONModel can guide designers to the potential source of the discrepancy, and, if appropriate, can semi-automatically evolve the models to handle unanticipated inputs.

KEYWORDS: Management, what-if, behavioral modeling, active probing