### ScaLAPACK band matrix or dense matrix LU?

Posted:

**Tue Feb 21, 2006 4:12 pm**Hello,

I have a very large general matrix (N = 87040) that has an upper band and a lower band of equal size (bw = 8703). Would this still be considered a "narrow" band? A compact block-column storage would save me a lot of space, but will I pay for it by using the divide and conquer algorithms in the LU for narrow band matrices in comparison to the dense matrix LU?

One other question, is the 1D block-column distribution described in the ScaLAPACK user guide intended to be cyclic? For example, if I have 121 processors for my matrix (see above) of NxN, do I use a block size equal to 720, such that the last processor has 640 columns. Or should I be using a smaller block of 200 and cycle through my 1D processor grid?

Thanks in advance

Cheers

Rob

I have a very large general matrix (N = 87040) that has an upper band and a lower band of equal size (bw = 8703). Would this still be considered a "narrow" band? A compact block-column storage would save me a lot of space, but will I pay for it by using the divide and conquer algorithms in the LU for narrow band matrices in comparison to the dense matrix LU?

One other question, is the 1D block-column distribution described in the ScaLAPACK user guide intended to be cyclic? For example, if I have 121 processors for my matrix (see above) of NxN, do I use a block size equal to 720, such that the last processor has 640 columns. Or should I be using a smaller block of 200 and cycle through my 1D processor grid?

Thanks in advance

Cheers

Rob