This project investigates the feasibility, cost, and potential impact of bounded-delay content distribution networks.  A PlanetLab prototype is used to experiment with content placement and delivery subject to end-to-end delay bounds.

We designed a distributed algorithm to dynamically select a subset of servers of a CDN to host the content such that a global delay bound is guaranteed.  Evaluation results drawn from our implementation and deployment demonstrate that despite Internet delay variability, subsecond delay bounds can be guaranteed with a very high probability at only a moderate content replication cost.

We further investigated a load-balancing issue in the system which necessitates considering servers' workload conditions and the popularities of content objects when deciding on replication strategies.

A generalized schedulability-maximizing problem was then formulated and investigated for large-scale distributed systems.  We explored a cluster of resource allocation policies, and identified promising candidates.  Important insights were provided into the problem for subsequent analysis.

Distributed systems continue to grow in scale and complexity, resulting in increasingly more involved interactions among components and increasingly more intricate failure modes that are very hard to diagnose manually. This increased vulnerability of larger systems, together with the increased difficulty of failure diagnosis, has motivated machine learning approaches to automate the diagnosis task. While preliminary encouraging results are achieved, scaling up the existing approaches to large applications remains challenging. With increase in scale, current approaches suffer the curse of dimensionality exacerbated by the exploding set of system states and measured metrics. In this project, we present techniques that significantly improve scalability of automated performance diagnosis methods.

We developed various protocols, including localization, multi-frequency MAC, and distributed storage systems for large-scale wireless sensor networks in which individual nodes subject to severe resource constraints.

Production run software failures cause endless grief to end-users, and endless challenges to programmers as they commonly have incomplete information about the bug, facing great hurdles to reproduce it. Users are often unable or unwilling to provide diagnostic information due to technical challenges and privacy concerns; even if the information is available, failure analysis is time-consuming.

We propose performing initial diagnosis automatically and at the end user?s site. The moment of failure is a valuable commodity programmers strive to reproduce -- leveraging it directly reduces diagnosis effort while simultaneously addressing privacy concerns. Additionally, we propose a failure diagnosis protocol. So far as we know, this is the first such automatic proto- col proposed for online diagnosis. By mimicing the steps a human programmer follows dissecting a failure, we deduce important failure information. Beyond online use, this can also reduce the effort of in-house testing.