Forward Thinking in Reverse

Developed during our recent six-month engagement at Dell—a leading computer manufacturer and services provider with one of the world's leading supply chains—we discuss a network flow-based mixed-integer linear program (MILP) model to identify the critical factors in optimizing reverse supply chain design decisions to optimize profit. The model is fast, intuitive, flexible, and robust to uncertainty. Its outputs include specific design recommendations, financial impact estimates, dynamically generated product routing diagrams, and multi-scenario sensitivity analysis. Through two case studies, the first in U.S. smartphone returns and the second in Europe, Middle East and Africa (EMEA) Alienware-branded computer returns, we show how our model fosters standardized, robust strategic decision-making and serves as a platform upon which to build management systems for continuous improvement. We then discuss two such systems: a simulation-based reusable packaging cost-benefit analysis (CBA) calculator, and an automated dashboard for managing disassembly-for-parts decisions.

MailTalk

With the goal of improving users' productivity, we use the Sphinx-4 speech recognition library to enhance Mac OS X's email client, Mail, with a speech-based command interface. We find that under Sphinx-4's baseline configuration, command accuracy (CAcc) varies from 89-99% in our controlled tests. After tuning, we improve accuracy by up to 3 percentage points, to a CAcc of 92-99%. We conclude that that while the productivity gains are not likely realized in this implementation, future work holds the promise of doing so. We also contribute a novel extension to the Sphinx-4 API to automate testing using WAV files as input.

Structuring Genomic Data in GEO Using Agglomerative Clustering

The majority of genomic data in the Gene Expression Omnibus (GEO) is not structured, making it diffuclt for researchers to use this data. We offer an adapted agglomerative clustering technique based on sample names to give structure to this data, and name the resultant clusters based on both itemset mining and acronym expansion. To visualize results, we use a proxy server implementation to "inject" the structured results into live GEO website results. We find these clusters look intuitively useful upon visual inspection, yet when compared with pre-computed clusters where available, we find the Rand index is typically below 0.5, indicating relative disagreement.

Reexamining Operating System Support for Database Management

In 1981, Michael Stonebraker observed that database management systems written for commodity operating systems could not effectively take advantage of key operating system services, such as buffer pool management and process scheduling, due to expensive overhead and lack of customizability. The “not quite right” fit between these kernel services and the demands of database systems forced database designers to work around such limitations or re-implement some kernel functionality in user mode.

We reconsider Stonebraker's 21-year old observations in the context of a modern-day database system, Microsoft SQL Server 2000, and the commodity operating system for which it is explicitly designed, Microsoft Windows 2000. We show that operating system services have become more efficient and flexible so as to meet many of SQL Server's needs directly. We also identify areas where operating system services continue fall short of meeting the needs of a DBMS, and propose several enhancements to rectify these shortcomings.