next up previous
Next: User Access Analysis Up: An Integrated Metric for Previous: The Workload Model

Empirical Workload Measurements and Analysis

   figure4284
Figure 2: VOSAIC Data

Our study is based on an understanding of real video workload environments, user behavioral patterns and system responses to changes in workload. We start by identifying a baseline class of workload components and a simplified characterization of the workload. For instance, we expect the video workloads to be largely I/O bound and not CPU bound gif.

Our metrics and measurements are derived from different Video-on-Demand (VOD) systems implemented at the University of Illinois, Urbana-Champaign. The first VOD system considered is the VOSAIC web system and we will observe the user access log characteristics obtained over a 20 day observation period of the VOSAIC (video-over-the-internet) system [4]. The user access logs allow us to analyze the user access statistics and general user information important for modeling of user satisfaction (see section 4.1). The second system considered is a hierarchical VOD system with a hierarchical set of video servers [15]. In this system, we observe performance characteristics such as jitter and frame loss rate. The third system is a simple remote VCR system with a single VOD server and client that we use to observe synchronization skew performance characteristics. The jitter, loss rate and skew characteristics allow us to analyze the system access and resource dependencies important for modeling of resource consumption (see section 4.2). Before we discuss performance analysis results, we briefly present the main architectural features of the three VOD systems.




next up previous
Next: User Access Analysis Up: An Integrated Metric for Previous: The Workload Model

Klara Nahrstedt
Fri Oct 3 16:05:57 CDT 1997