SC13 Home > SC13 Schedule > SC13 Presentation - Asynchronous PDE Solver for Computing at Extreme Scales

SCHEDULE: NOV 16-22, 2013

When viewing the Technical Program schedule, on the far righthand side is a column labeled "PLANNER." Use this planner to build your own schedule. Once you select an event and want to add it to your personal schedule, just click on the calendar icon of your choice (outlook calendar, ical calendar or google calendar) and that event will be stored there. As you select events in this manner, you will have your own schedule to guide you through the week.

Asynchronous PDE Solver for Computing at Extreme Scales

SESSION: Research Poster Reception

EVENT TYPE: Posters, Electronic Posters, and Education Posters

TIME: 5:15PM - 7:00PM

AUTHOR(S):Konduri Aditya, Diego Donzis, Torsten Hoefler

ROOM:Mile High Pre-Function

Scalability of codes at extreme scales is a critical issue of current relevance. At extreme levels of parallelisms, communication between processing elements(PEs) could represent a substantial running time, resulting in substantial waste in computing cycles. We investigate a novel approach based on common finite-difference schemes for partial differential equations(PDEs) in which computations are done locally with values already available at each PE, without waiting for updated data from neighboring PEs. No synchronization among cores is enforced and computations proceed regardless of messages status. This drastically reduces processor idle times, improving computation rates and scalability. We show that accuracy of common numerical schemes is reduced when asynchronicity is present. We derive new schemes that can maintain the accuracy under asynchronous conditions. These new schemes are implemented and tested. Performance is evaluated through statistics of different measures of asynchronicity. A new implementation of RDMA communications is shown to provide significant advantages .

Chair/Author Details:

Konduri Aditya - Texas A&M University

Diego Donzis - Texas A&M University

Torsten Hoefler - ETH Zurich

Add to iCal  Click here to download .ics calendar file

Add to Outlook  Click here to download .vcs calendar file

Add to Google Calendarss  Click here to add event to your Google Calendar