Distributed packed storage format
Definition :
ScaLAPACK 2-D block cyclic distribution (elementary block of size s)
process griddistributed packed format :
matrix is partitioned into square blocks of size b such that
(distributed blocks)implementation based on ScaLAPACK or PBLAS routines
a unique ScaLAPACK array for the packed structure
Implementation
Example :
process grid,
elementary block 0
grid block 0
1
2
3
4
5
distributed block 0
1
2
0
1
2
3
4
5
3
4
5
0
1
2
0
1
2
3
4
5
3
4
5
0
1
2
0
1
2
3
4
5
3
4
5
Tuning parameters
determine the dominant operation (e.g DGEMM in a sequential Cholesky)
choose ScaLAPACK parameters for the dominant operation:
optimize elementary block size
optimize
process grid
choose block size
:that optimizes the dominant operation
consistant with an accepted extra-storage due to diagonal blocks (tradeoff)
