You are in the main area:Events
Headerimage for: Workshops


Cray GPU workshop

25. 10. 10
Cray Workshop on Hybrid Manycore Programming and GPU Accelerators in High Performance Computing

HLRS, Room 1.067, Nobelstr. 19, 70569 Stuttgart.

2010, Monday Oct. 25, 9:30 - 15:30

Abstract: Moving to Exascale computing will require significant hardware innovation. To achieve an Exascale within the power budget proposed by potential Exaflop customers, hardware developers need to supply something different. Many feel that the number of nodes in the largest HPC system will remain fairly constant; however, the nodes themselves will increase in performance by factors of 500-1000. By adding accelerators and/or more less power hungry cores to the node, performance can increase without a significant increase in power consumption. These designs put a tremendous burden on the application programmer. Currently most applications use MPI entirely, few employ any type of shared memory threading. These applications will have to change to effectively utilize future HPC architectures.

This Hybrid Multicore presentation will cover the issues the application programmers face and what can be done now to prepare applications for these future high performance nodes. Cray is working on a Hybrid Multicore programming environment to supply tools to facilitate the porting and optimization of HPC applications. These tools will not be automatic, as no one in the industry really believes that a compilation system can perform the analysis and transformations to automagically transform the application. What the tools do is converse with the user through a user friendly interface to examine the entire application. The programming environment uses a database representation of the applications, giving it the ability to perform whole program analysis. Given the capabilities of the system and the ability for the application developer to direct the analysis, we believe that the programming environment will significantly reduce the time required to transform the application into a form that will be able to use the more powerful nodes.
The intent of the session will be to gather feedback from the attendees on capabilities that might be missing in the programming environment. Delivery of the programming environment system is targeted for first quarter 2012; however, pieces of the system will be available later this year. There is still time for the user community to have an impact on the design and ultimate capabilities of the system.

Further information and registration can be found here.