COMPUTATIONAL AND DATA SCIENCES (CDS)
CDS Research Seminar
SPEAKER : Dr. Konduri Aditya, Sandia National Laboratories, Livermore, USA.
TITLE: Highly scalable asynchronous algorithm for partial differential equations: a path towards Exascale
DATE: Wednesday, September 20, 2017
TIME: 11:00 AM
VENUE: Room 102, CDS Seminar Hall (First Floor)
ALL ARE CORDIALLY INVITED
Synchronization overheads pose a major challenge as applications advance towards extreme scales. In current algorithms for partial differential equations (PDEs), synchronization as well as data communication delay the parallel computations at each time step and may create a new scaling wall when moving towards Exascale. In this work, we present an asynchronous computing algorithm based on finite difference schemes for PDEs where no synchronization between processing elements (PEs) is enforced. PEs are allowed to continue computations regardless of messages status and are thus asynchronous. We show that accuracy of commonly used finite difference schemes is degraded when they are used asynchronously. Since message arrivals at PEs are essentially a random process, so is the behavior of the error. Within a statistical framework we show that average errors drop always to first-order regardless of the original scheme. The value of the error is found to depend on both grid spacing as well as characteristics of the computing system including number of PEs and statistics of the delays. We propose new schemes that are robust to asynchrony and present a novel asynchronous algorithm for their implementation. We utilize modern remote memory access programming schemes that have been shown to provide significant speedup on modern supercomputers, to efficiently implement communications suitable for asynchronous schemes. We present results from simulations that demonstrate excellent scalability of the new asynchronous schemes for large-scale computing.
Biography of the Speaker:
Dr. Aditya is a gold medalist in bachelors program of Mechanical Engineering from Osmania University, Hyderabad, which he completed in 2007. He obtained Masters degrees from JNCASR and Georgiatech in 2009 and 2011 respectively. He obtained his PhD degree in Aerospace engineering from Texas A&M University, College Station, Texas in 2015, where he was awarded Outstanding dissertation award. From 2016 onwards, he has been working as a Postdoctoral Research Associate at Combustion Research Facility, Sandia National Laboratories, Livermore, California, USA. To his credit, he also won Best Graduate Student Presentation Award, American Physical Society meeting – Texas 2014. His Research interests include numerical methods for high performance computing, machine learning and computational fluid dynamics. Currently, he is working on simulations of combustion phenomena at scale, asynchronous computing algorithms for solving partial differential equations at extreme scales and in-situ data analytics using machine learning.