Best practices for building hardware designs for living computational science applications.

摘要

Scientific computing or Computational science, is a field of study where engineers and scientists use computer simulations to solve equations that model the physical world. In some cases, these equations come from the first principles of physics. In the past, these simulations were run on a single processor machine. However, due to various technological reasons, the performance of these machines are not likely to improve at the same rate as in the past. In order to improve the performance per watt of these simulations, special-purpose hardware accelerators can be used. This work mainly focuses on using FPGA-based hardware accelerators. In order to run these simulations on an FPGA accelerator, the application code needs to be re-factored into software and hardware sections. These faster simulations have motivated scientists to capture more behavior of the physical world. As additional behavior is captured, the application code needs to be re-factored each time, and a significant effort is required to re-build the design. Unfortunately, these multiple cycles of re-design reduces the overall productivity of scientists and engineers.;This work proposes a set of hardware design guidelines for changing computational science codes or living computational science codes. These guidelines co-evolve the hardware with the software, reducing the overall effort of re-design and improving productivity. The design guidelines are evaluated for effectiveness, communicability, and broad applicability. Experimental results have shown that the overall re-design effort is reduced, and these guidelines are broadly applicable to a wide variety of scientific computing applications.