Recent Presentations
Scheduling to balance energy and delay
No longer is faster always better in system design. Nowadays, across all levels of computer systems, speed costs power and power costs money -- so performance must be balanced with energy usage. The most common approach for balancing energy consumption and performance is dynamic speed scaling, which adapts the processing speed to the current workload. The focus of this talk is to understand some fundamental questions about speed scaling, such as: What are the optimal speeds? How do the optimal speeds depend on the scheduling of the system? What improvement does dynamic speed scaling provide over simple schemes such as "sleep when idle"?
[ppt]The impact of local scheduling in load balancing designs
How far is load balancing from optimal, and how does the answer depend on the local scheduler? How much can be gained by changing the local scheduler in load balancing designs?
[slides]-
Does helping the little guy help everyone?
A five hour survey of recent work studying the performance of policies that favor small jobs. Does favoring small jobs hurt large jobs? How much can be gained? Can such policies be utilized in practice?
[part1] [part2] Applying to Ph.D. programs in CS
An overview of the process of applying to grad school.
[slides]Scheduling with inexact job size information
How accurate must job size information be to achieve a desired quality of service? This talk provides new analytic results to help designers answer this question.
[slides]Some lessons scheduling and queueing can teach us about system design
This talk provides a overview of some basic and some not-so-basic results from scheduling and queueing and shows how they can be useful for system designers. The talk is meant to be light, interactive, and fun, but to still provide a feeling for the type of research questions I like to approach.
[slides]Revisiting the performance of large jobs
In recent years, the response times experienced by large job sizes have been the focus of a growing number of papers. Though results about many scheduling disciplines have appeared, to this point, results characterizing the response time of large job sizes have been limited to either mean value analysis or law of large numbers scalings. This talk presents a novel framework that unifies these results and provides new results characterizing the distributional behavior of large job sizes.
[slides]Fairness in queues
The growing trend in computer systems towards using scheduling policies that prioritize jobs with small service requirements has resulted in a new focus on the fairness of such policies. In particular, researchers have been interested in whether biasing towards small job sizes results in large jobs being treated "unfairly." However, unfairness is an amorphous concept and thus difficult to define and study. In this talk, I will present some recent attempts to define and study the concept of fairness in single server queueing settings.
[slides]Open versus closed: A cautionary tale
Workload generators may be classified as based on a closed system model, where new job arrivals are only triggered by job completions (followed by think time), or an open system model, where new jobs arrive independently of job completions. In general, system designers pay little attention to whether a workload generator is closed or open. In this talk, I will illustrate (using a combination of implementation and simulation experiments) that there is a vast difference in behavior between these open and closed models in real-world settings.
[slides]A unified framework for modeling TCP Vegas, TCP SACK, and TCP Reno
We present a general analytical framework for the modeling and analysis of TCP variations. The framework allows the modeling of multiple variations of TCP, including TCP-Vegas, TCP-SACK, and TCP-Reno, under general network situations.
[slides]-
Dimensionality reduction: 4 example applications
We present an approach for solving 2 dimensionally infinite Markov chains, which we refer to as dimensionality reduction. The technique transforms a chain from a 2D-infinite chain into a 1D-infinite chain which can be analyzed. To illustrate the applicability of this technique, in this sequence of talks we apply the dimensionality reduction technique to analyze several variants of the cycle stealing problem as well as several variants of multiserver priority systems.
[part1] [part2] [part3] [part4]