Throughput-Buffering Trade-Off Analysis for Scenario-Aware Dataflow Models

Abstract

In multi-media applications, buffers represent storage spaces that are used to store the data communicated between different tasks in the application, and throughput refers to the rate at which output data is produced by the application. The capacities of the buffers influence the throughput, by altering the waiting times for tasks that need to read or write data from or to the buffers. The buffers are realized using memory. To minimize the memory usage, we look for algorithms to compute the minimal capacity requirements for buffers to execute an application under a given throughput constraint. Synchronous dataflow (SDF) is a common formalism used to model applications in such algorithms. SDF however, is not suitable to describe today's dynamic applications, as it cannot express task variations. Finite-State-Machine Scenario-Aware Dataflow (FSM-SADF) is an extension of SDF that allows for not only task variations, but also structural variations, making it suitable for a wide range of dynamic applications. This paper provides the first throughput-bufering trade-of analysis for FSM-SADF models. The analysis provides the Pareto space of throughput and storage space trade-offs. The trade-off analysis is done by a guided Design Space Exploration (DSE) that cuts-off the exploration on non-critical buffers. The core of such a DSE is an FSM-SADF throughput analysis that, given the capacity of every buffer, obtains the throughput, as well as the critical buffers. We demonstrate the feasibility of our analysis with a number of examples.