Performance Estimation for the Codesign of Systems Running Concurrent Applications

Asawaree Kalavade
Bell Labs

Thursday, April 24th, 1997
Hogan Room, 531 Cory Hall
5:00-6:00 p.m.


Network-centric embedded systems such as PDA's and settop boxes are growing in popularity. Such systems are expected, in their next generation, to support concurrent sophisticated applications like audio, video, and graphics. These applications offer an attractive feature in that they can adapt to the network conditions. For example, the frame rate of a video encoder (and hence its processing demands) may vary with network congestion. Similarly, one of several audio algorithms (PCM, ADPCM, or LPC) may be selected at run time, depending on the required output bitrate. Such adaptations are typically controlled by the feedback from real-time protocols like RTCP. The downside of such adaptive applications is that designing embedded systems to support them is much harder because performance estimation becomes nontrivial. The reason for the latter is that the processing load on the system is a complex dynamic process that is a function of several varying phenomena -- (1) dynamically changing adaptations of each application, (2) the variable execution times of different tasks within each application, and (3) the relative timing between the initiation of different applications.

In this work, we address the performance estimation issue. Our objective is to codesign an embedded system architecture that runs a set of such adaptive applications. We model each application as a task graph, where a particular adaptation corresponds to a particular execution path through this graph. We present a theoretical framework to analyze the state of computation of the system. We compute the exact processing delay distribution for each adaptive application. We also compute other performance metrics such as average nodal wait and average resource wait. These performance measures are then used to design and verify the embedded systems.

This work is being done jointly with Pratyush Moghe, Bell Labs.