David Finkel

Thin to win? Network performance analysis of the OnLive thin client game system

The growth in network bitrates and server-based processing has provided a renewed opportunity for thin client games, where the server does heavy-weight computations, sending only the visual game frames to the client, and the client displays frames, sending only the user actions to the server. Understanding the traffic characteristics of thin client games is important for building traffic models and classifiers and planning network infrastructures to meet future demand. This paper provides the first detailed study of the network characteristics of OnLive, a commercially available thin client game system. Carefully designed experiments measure OnLive game traffic for several game genres, analyzing the bitrates, packet sizes and inter-packet times for both upstream and downstream game traffic, with comparisons to traditional game clients and streaming video. Results indicate OnLive rapidly sends large packets downstream, similar but still significantly different than live video. OnLive less frequently sends much smaller packets upstream, significantly different than traditional game client traffic. The results should be a useful beginning for building effective traffic models and classifiers, and for preparing end-host networks to support this upcoming generation of computer games.

The effects of latency on player performance in cloud-based games

Cloud-based games are an increasingly popular method to distribute and play computer games on the Internet. While there has been some work studying network aspects of cloud-based games and examining the effects of latency on traditional games, there has not been sufficient research on the impact of latency on cloud-based games nor a comparison of the impact of latency on cloud-based games versus traditional games. This paper presents the results of two user studies that measure the objective and subjective effects of latency on cloud-based games, one study using the commercial cloud game system OnLive and the other study using the academic cloud game system GamingAnywhere. Analysis of the results shows both quality of experience and user performance degrade linearly with an increase in latency. More significantly, latency affects cloud-based games in a manner most similar to that of traditional first-person avatar games, the most sensitive class of games, despite the fact that the cloud-based games may have a different user perspective. These results have implications for cloud-based game designers and cloud system developers.

Peer learning in an introductory computer science course

A problem in teaching large introductory computer science courses is to overcome the impersonality of the large lecture class and to provide more personal attention to individual students. Our approach is to use peer learning experiences to instill in students the need to take responsibility for their learning and for the learning of those around them. Recent work has shown that educational quality for students and productivity for faculty can be enhanced through use of peer-learning environments where students do not just learn and faculty do not just teach. The novel aspects of our work are to apply group learning in a large introductory computer science class setting and to expect more responsibility on the part of students for their learning. In support of these goals we have introduced the use of upper-level undergraduate students to help facilitate student group interaction. In addition, we have developed software to minimize the administrative overhead of handling many groups and for students to electronically record group learning activities.

On the performance of OnLive thin client games

Computer games stand to benefit from “cloud” technology by doing heavy-weight, graphics-intensive computations at the server, sending only the visual game frames down to a thin client, with the client sending only the player actions upstream to the server. However, computer games tend to be graphically intense with fast-paced user actions necessitating bitrates and update frequencies that may stress end-host networks. Understanding the traffic characteristics of thin client games is important for building traffic models and traffic classifiers, as well as adequately planning network infrastructures to meet future demand. While there have been numerous studies detailing online game traffic and streaming video traffic, this paper provides the first detailed study of the network characteristic of OnLive, a commercially available thin client game system. Carefully designed experiments measure OnLive game traffic for several game genres, analyzing the bitrates, packet sizes and inter-packet times for both upstream and downstream game traffic, and analyzing frame rates for the games. Results indicate OnLive rapidly sends large packets downstream, similar but still significantly different than live video. Upstream, OnLive less frequently sends much smaller packets, significantly different than upstream traditional game client traffic. OnLive supports only the top frame rates with high-capacity end-host connections, but provides good frame rates with moderate end-host connections. The results should be a useful beginning to building effective traffic models and traffic classifiers and for preparing end-host networks to support this upcoming generation of computer games.

Autonomic cluster management system (ACMS): a demonstration of autonomic principles at work

Cluster computing, whereby a large number of simple processors or nodes are combined together to apparently function as a single powerful computer, has emerged as a research area in its own right. The approach offers a relatively inexpensive means of achieving significant computational capabilities for high-performance computing applications, while simultaneously affording the ability to increase that capability simply by adding more (inexpensive) processors. However, the task of manually managing and configuring a cluster quickly becomes impossible as the cluster grows in size. Autonomic computing is a relatively new approach to managing complex systems that can potentially solve many of the problems inherent in cluster management. We describe the development of a prototype automatic cluster management system (ACMS) that exploits autonomic properties in automating cluster management.

An open source laboratory for operating systems projects

Typical undergraduate operating systems projects use services provided by an operating system via system calls or develop code in a simulated operating system. With the increasing popularity of operating systems with open source code such as Linux, there are untapped possibilities for operating systems projects to modify real operating system code. We present the hardware and software configuration of an open source laboratory that promises to provide students that use it with a better understanding of operating system internals than is typically gained in a traditional operating systems course. Our preliminary projects and evaluation suggest that thus far the lab has achieved its primary goal in that students that used the lab feel more knowledgeable in operating systems and more confident in their ability to write and modify operating system code.

Improving the Productivity of Volunteer Computing by Using the Most Effective Task Retrieval Policies

Volunteer computing projects have been used to make significant advances in knowledge since the 1990s. These projects use idle CPU cycles donated by people to solve computationally intensive problems in medicine, the sciences and other disciplines. It is important to use the donated cycles as efficiently as possible because participation in volunteer computing is low and the number of volunteer computing projects keeps increasing. Task retrieval policies, policies describing when a volunteered computer requests additional work from a server, can have an effect on the number of wasted CPU cycles and consequently, the number of tasks completed by clients. We present the results of simulating different task retrieval policies for clients under realistic conditions, including clients running on computers with one single-core CPU, clients running on computers with multi-core CPUs, and clients running on computers that are put into a power save mode by environmentally conscious owners.

Load balancing in a multi-server queuing system

Abstract A queuing system consists of N identical, exponential service channels, each with its own Poisson stream of customers. If allowed to run unattended, the system may become unbalanced; that is, some queues may become long while others are empty. In this case, it may be beneficial to transfer jobs from the long queues to shorter ones to “balance the load”. We assume that the system incurs a cost for each customer waiting in the queue and a cost each time a job is transferred. We show that, in the case of light traffic intensity, it is optimal to transfer a job to the shortest queue whenever the number of jobs in its own queue exceeds the number of jobs in the shortest queue by a fixed threshold. We also include simulations to test how well this policy performs under moderate and heavy traffic intensities and to compare this policy to other reasonable, implemcntable policies.

Webware: a course about the Web

Sophisticated applications and software development on the Web demand an extensive and thorough understanding of a variety of computer science disciplines, as well as providing their own set of issues. Therefore, we have created an advanced undergraduate computer science course called Webware: Computational Technology for Network Information Systems that builds upon and extends knowledge previously gathered by the students. We describe its contents, our teaching experience, and address the challenges of teaching both the foundations and current technological issues of Web programming.

Experience With Peer Learning in an Introductory Computer Science Course

A problem in teaching large introductory computer science courses is to overcome the impersonality of the large lecture class and to provide more personal attention to individual students. We used peer learning experiences to increase student interaction in a large introductory class and to instill in students the need to take responsibility for their learning and that of those around them. In support of this approach we have introduced the use of upper‐level undergraduate students to help facilitate student group interaction and developed software to minimize the administrative overhead of handling many groups. The article describes our work and reports on results, which show that responsibility for learning was transferred from the instructor to the students. Peer learning was most accepted by students who reported having trouble with the course while the approach had mixed support among students who did well in the course. Our overall assessment, based on two offerings of the course with emphasis on gr...