[Updated on Apr. 18] Integration of CiNii Articles into CiNii Research


Bibliographic Information

Other Title
  • FLAT: An MPI Friendly GPGPU Programming Framework for GPU Clusters

Search this article



A program for a PC cluster which equips GPUs consists of two types of code, for GPUs and for CPUs. The GPU code executes parallelized algorithms to introduce high speed computing supported by a CPU code which performs communication with other nodes. Although MPI library is commonly utilized to transfer data in the CPU code, MPI functions can not be written in the GPU code. Programmers are forced to implement CPU and GPU codes alternately with taking care of data movement among nodes. In order to reduce software development costs, we propose a programming framework called FLAT which enables GPU codes to embed MPI functions. This paper describes execution model and implementation of FLAT, and discusses availability and performance obtained by two case studies, Livermore Loop18 and optical flow programs. Through the experimental results, we confirmed that FLAT increases readability in synthesized GPU codes with maintaining bearable performance degradation, which is less than 3% for a coarse-grained parallel program.


Citations (0)*help

See more


See more

Related Articles

See more

Related Data

See more

Related Books

See more

Related Dissertations

See more

Related Projects

See more

Related Products

See more


Report a problem

Back to top