High performance computing (HPC) has become a critical tool for industries seeking solutions to complex engineering and physical problems. While the problems and underlying governing equations describing the transport physics have not changed in decades, higher fidelity simulations have pushed the boundaries of computational science. The machines employed for computations have undergone a radical transformation. In the previous decades, computing was dominated by single-core hardware. Each generation of silicon process technology offered greater clock speeds, and hence, scientific codes which were mostly compute bound, simply ran faster without any modifications. However, power constraints, especially heat dissipation, led to a plateauing of clock speeds around the mid 2000’s. Consequently, with the advent of modern many-core and multi-core processors such as GPUs, the primary way to achieve performance gains has been through parallelism. A large and diverse collection of scientific software applications that are widely used in academia, national labs, and industry now face enormous challenges in catching up with this paradigm shift in the hardware landscape. Today’s HPC practitioners are faced not only with complex multi-level parallelism but also massive parallelism available on architectures such as GPUs. This presents both challenges and great opportunity. In this talk, I will discuss some of the unique algorithmic, computational and engineering challenges involved in developing an engineering simulator by efficiently utilizing the parallelism offered by GPUs. As an example, one computational challenge is to consider parallelism from the earliest stages of application design for maximal performance and not as an afterthought as is done more commonly. Performance results from high fidelity reservoir simulations of synthetic and real-field assets will also be presented.
Speaker: Karthik Mukundakrishnan, Stone Ridge Technology
Contact:Website: Click to Visit
Save this Event:iCalendar
Windows Live Calendar
Share this Event:Email to a Friend
Stanford, CA 94305
Website: Click to Visit