On the detailed design issues of a multi-agent architecture based on CUDA threads towards efficiently Simulating WSN Systems

Emulators used in sensor network simulation are bounded to specific hardware. Adding different modules of hardware (RF´s, MCU´s, FPGA´s) to these simulators-emulators is a hard and time consuming task. In this paper we extend an integrated generic emulator, module of a recently proposed, by the authors [3], WSN simulator architecture, capable of emulating any MCU/FPGA/DSP chip functionality completely described in XML file format with integrating CUDA emulated capabilities. The suggested architecture is based on novel Multi-agent and Multi-Threading technology concepts and is designed for the CUDA ready implementation. The basic modules and the pseudocode of this implementation is herein presented along with the essential setup. In [3] we proposed the agent mechanism (token interpreter) as the backbone of a WSN simulator. In this paper we extend in detail the design for the use of this interpreter as a CUDA thread, dealing with problems such as communication cost, thread conflict and kernel concurrent execution. We analyze the implementation of data structures, we introduce the necessary transformations to improve performance as well as we present initial validation of its performance regarding multithreading overheads for the CPU and GPU implementations.