The FERMAT Lab performs research in the area of embedded system design for critical applications.
Souping up hardware sim speed
Sandeep Shukla has led a team to improve the simulation performance of hardware models created in a language called SystemC. The team’s new infrastructure, SCGPSim, has garnered a best paper award at the 15th Asia and South Pacific Design Automation Conference (ASP-DAC) in Taiwan.
Shukla and his collaborators said that they were able to demonstrate how to speed up the simulation performance of certain SystemC-based hardware models “by exploiting the high degree of parallelism afforded by today’s general purpose graphic processor units (GPGPU).” These units have multiple core processors capable of very high computation and data throughput. When parallelism is applied, it means that the processor units can run various parts of the simulations simultaneously, and not just as a single sequence of computations. Their experiments were carried out on an NVIDIA Tesla 870 with 256 processing cores donated to the lab by NVIDIA.
In the past, Shukla said, “significant effort was aimed at improving the performance of SystemC simulations, but little had been directed at making them operate in parallel. And none of the attempts were ever targeted at a massively parallel platform such as a general purpose graphic processor unit.” Shukla said their preliminary experiments showed they were able to speed up SystemC based simulation by factors of 30 to 100 times that of previous performances.
Another aspect of their work was the use of a specific programming model called Compute Unified Device Architecture (CUDA). It is an extension to the C software language that “exploits the processing power of graphic processor units to solve complex compute-intensive problems efficiently,” Shukla explained. “High performance is achieved by launching a number of threads and making each thread execute a part of the application in parallel.”
The Air Force Office of Scientific Research and the National Science Foundation helped support this research.
— Lynn Nystrom