Dynamically-Tunable Dataflow Architectures Based on Markov Queuing Models

Dataflow architectures are fundamental to achieve high performance in data-intensive applications. They must be optimized to elaborate input data arriving at an expected rate, which is not always constant. While worst-case designs can significantly increase hardware resources, more optimistic solutions fail to sustain execution phases with high throughput, leading to system congestion or even computational errors. We present an architecture to monitor and control dataflow architectures that leverage approximate variants to trade off accuracy and latency of the computational processes. Our microarchitecture features online prediction based on queuing models to estimate the response time of the system and select the proper variant to meet the target throughput, enabling the creation of dynamically-tunable systems.