Slow ScaLAPACK performance

[Pete]

Hello,

how fast can you solve a SVD for a 1000x1000 matrix? On my Dual Core AMD Opteron (Linux, 64-bit) it takes more than a minute. And for a 2000x2000 matrix it takes more than 10 minutes. Is that normal?

[Julien Langou]

So first are you using LAPACK or ScaLAPACK?

For LAPACK, it depends on your machine and your BLAS, it can depend
also on your input matrix (but less). For random 1000-by-1000 matrix,
I got the svd in about 22.9 sec. For random 2000-by-2000 matrix, I got the
svd in about 204.7 sec.

If you go from 1000-by-1000 to 2000-by-2000, the time to solution is (very
roughly) multiplied by 8 since the cost of the SVD is O(n^3). So going from
1 minute to 10 minutes seems about right in your case. In my case, I go
from 22.9 sec to 204.7 sec. Well almost a factor of 8.

Best wishes,
julien