Overview

My research area is systems. Specifically I am interested in distributed systems, storage management, mobility, and adaptive systems. Recurring themes in my work are performance optimization and autonomy.

Systems research is analogous to the development and tuning of a high-performance engine. The tuning aspect of this analogy can be described as a quest for balance, stability and equity. In storage management and file systems we may sacrifice capacity and simplicity of metadata management for decreased average user response time and increased reliability. The best balance is defined by the objectives of the system. The system's researcher attempts to develop techniques that improve the required performance in light of changes in technology.

The other theme in my research is the quest for autonomy and self-regulation. In my work this involves the search for self-tuning or self-regulating systems algorithms that require little or no user or administrator input. If systems research is analogous to designing and tuning a high-performance engine (performance can be measured as power, speed, reliability, mileage, simplicity of repair etc.), and if a recent goal is increasing the autonomy of computer systems, then my research goal is analogous to the desire to construct engines that are capable of repairing themeselves, as well as retuning themselves in response to environmental and usage changes. An advantage of adaptive and on-line algorithms (from the A.I. and machine learning domain, and seeing growing adoption into advanced systems efforts) is that they often have such self-regulation as an inherent feature. The adaptive modification of a mobile disk's spin-down time-out is an excellent example of such algorithms.

Predictive Grouping - My PhD dissertation looks into the automated on-line grouping of data. Data access patterns are dynamically observed, and groups of related items are built based on the relative strengths of pair-wise file relationships. Such groupings are applied for improving distributed cache performance, and automating the integration of tertiary and advanced storage devices.

Noah: Successor Prediction - The prediction of the "next access event" is a direct off-shoot of my work on data grouping. Establishing a sound successor prediction strategy is complimentary to the automated determination of inter-file relationships. Successor prediction can be seen as providing likely pairings, and the Noah project aims to develop such successor prediction mechanisms that use minimal metadata while providing accuracy comparable to much more elaborate predictive schemes. A key feature of Noah is that it allows for adjustable accuracy, i.e., requiring a specific levels of confidence in the offered predictions. More recent work, in collaboration with J.F.Paris and Randal Burns has resulted in improved predictors also suitable for identifying most likely successors.

Mobile Hoarding - A persistent problem ..... This is a current USENIX-funded project aimed at developing improved data hoarding algorithms for mobile applications. It builds on the automated grouping work to provide a group-based mechanism for hoarding construction and better mobile file caching.

Predictive Power Conservation - In my first year at UCSC, I worked with Professors Helmbold and Long on a project for predictive power conservation in mobile applications. I worked on adapting the machine learning algorithm used to automatically adjust itself to optimal operating parameters. More details can be found on the old project web-page.

Workload Visualization - A recently launched project, in collaboration with the Visualization Group, attempts to provide a visual representation of file and data access behaviour. More details can be found on the project web-page.

Autonomy: Self-Tuning Systems - Too many controls, paramters, switches and knobs are an all too common result of building more elaborate and higher-performance systems. My work aims to provide autonomous solutions, requiring little or no paramter adjustment. Ideally, any parameter required by the system should be simple and meaningful to the end-user. Complex system configurations, and tricky administration are a common cuplrit for poorly-performing systems. Our most recent efforts in this area involve improved caching algorithms, that have demonstrated good performance under what would normally be adverse conditions for more common schemes.