So we have a 6,000,000 x 6,000,000 matrix of which only 130,000,000 million are NNZ (none zeros) which means only 0.002167% of the matrix is non zeros.
The above matrix needs to be multiplied to a vector containing 6,000,000 entries all of which are non zero.
how would you partition and how would you represent Blocks inside this matrix in order to achieve matrix multiplication using CUDA without utilizing hard disk swapping
P.S System details 12GB main memory and GTX260 graphics card
They are represented on disk in this manner
1: 2 3
3: 1 2 5
4: 5 6
5: 4 6
were the initial value represent the Node ID and anything behind the “:” represents out links to other nodes the above data structure forms a sparse matrix
(P.S Node 2 contains no out links therefore it is omitted from the file)
P.S i already tried partitioning the matrix using Ellpack R as specified in this paper but the data structure was using more than the main memory i have therefore wasn’t a feasible solution.
please do express anything you might feel is relevant since i am honestly stuck and need urgent help