Speaker: Dushyanth Narayanan, Carnegie Mellon University
Date: October 22, 1998
Reconfiguring large storage systems: the Argo migration planner
Large storage systems are difficult to configure and manage. Imagine a number of heterogenous devices, connected by a high-bandwidth network, that store data objects of varying sizes and access patterns and support clients with widely different performance requirements. Continually tweaking and tuning the parameters of this system for optimal performance is complex and non-intuitive, and extremely prone to human error.
Palladio is an architecture that proposes to automate these configuration and management tasks. I shall briefly describe Palladio, and the Forum configuration tool that it uses to generate optimal assignments (of data objects to storage devices). This optimal assignment can change with time, for various reasons -- changes in access patterns, disk failures, or addition of new devices.
Thus, we need a way to move the system from its current configuration to the optimal one by migrating and/or replicating data objects. While Palladio will provide support for these operations at the single object level, we need to be able to plan an entire sequence of these operations, to transition the system into its new state transparently and without downtime or loss of performance.
I shall describe the Argo migration planner, the plan generation heuristic that it uses, and some preliminary evaluations of the heuristic using synthetically generated workloads. Finally, I shall point out the important shortcomings in the current version of Argo, and identify directions for future research.