Heterogeneous cluster for acceleration of linear algebra computations

Authors: Šimeček I., Buk Z.


Keywords
GPGPU; grid computing; remote calls; library for numerical linear algebra.


Abstract
This project deals with a new concept of the heterogeneous grid for acceleration of the numerical linear algebra computing. We design this grid with respect to maximal ratio between cost and computational power. It allows a parallelization of scientific codes with minimal programming effort. We also optimize grid concept to be less sensitive to network parameters.


What's new
March 10, 2011 -  version 0.1 of this project is released.
July 20, 2011 -  version 0.2 of this project is released.
December 1, 2011 -  version 1.0 of this project is released.

Download:
source codes
Only for demonstartion purposes !
For fully funcionality it requires with:
1) CUDA SDK with CUBLAS ( for GPGPU dense matrix computing),
2) MPI implementation



Documentation:
User's guide PDF file (in Czech)
Final report PDF file (in Czech)



TODO list
Integration with:
1) CUSPARSE ( for GPGPU sparse matrix computing),
2) standard library for vector and matrix manipulation like ATLAS or GOTO,
3) sparse matrix solver (like SuperLU or UMFpack).


Used in papers:
APLIMAT 2012: Heterogeneous cluster for acceleration of linear algebra computations
ICPM 2011: Experimental grid for numerical linear algebra
SNA 2011: Experimental grid for numerical linear algebra
BACK