Performance Modeling of Virtualized Custom Logic Computations

Virtualization of custom logic computations (i.e., by sharing a fixed function across distinct data streams) provides a means of reusing hardware resources, particularly when resources are limited. This is common practice in traditional processors where more than one user can share processor resources. In this paper, we virtualize a custom logic block using C-slow techniques to support fine-grain context-switching. We then develop and present an analytic model for several performance measures (throughput, latency, input queue occupancy) for both fine-grained and coarse-grained context switching (to a secondary memory). Next, we calibrate the analytic performance model with empirical measurements. We then validate the model via discrete-event simulation and use the model to predict the performance and develop optimal schedules for virtualized logic computations. We present results for a Taylor series expansion of a cosine function with added feedback and an AES encryption cipher