c*******h
发帖数: 1096 | 1 any (including parallel computing) software can compute
partail svd on a matrix of size 200,000 x 600,000?
300 largest singular values are enough |
s****y
发帖数: 2052 | 2 你在umn？
lanczos行么？
【在 c*******h 的大作中提到】
: any (including parallel computing) software can compute
: partail svd on a matrix of size 200,000 x 600,000?
: 300 largest singular values are enough
|
l*****i
发帖数: 3929 | 3 lapack? not completely sure
【在 c*******h 的大作中提到】
: any (including parallel computing) software can compute
: partail svd on a matrix of size 200,000 x 600,000?
: 300 largest singular values are enough
|
c*******h
发帖数: 1096 | 4 yes i am at umn
i coded a lanczos version of svds in matlab, which uses much less memory
and runs much faster than the matlab function, but still it uses 4G memory
and 3G swap memory and runs for ever. After monitering the program for
4 hours, i gave up.
【在 s****y 的大作中提到】
: 你在umn？
: lanczos行么？
|
s****y
发帖数: 2052 | 5 if you are in umn, and are doing this kindo computing,
probably you can try supercomputing center. they have very good machines.
【在 c*******h 的大作中提到】
: yes i am at umn
: i coded a lanczos version of svds in matlab, which uses much less memory
: and runs much faster than the matlab function, but still it uses 4G memory
: and 3G swap memory and runs for ever. After monitering the program for
: 4 hours, i gave up.
|
c*******h
发帖数: 1096 | 6 to use msi machines is not a problem..
the problem is what software i can use..
i do not want to code a parallel version of svd using c from scratch..
【在 s****y 的大作中提到】
: if you are in umn, and are doing this kindo computing,
: probably you can try supercomputing center. they have very good machines.
|
s****y
发帖数: 2052 | 7 they have more memory。 how about that？
【在 c*******h 的大作中提到】
: to use msi machines is not a problem..
: the problem is what software i can use..
: i do not want to code a parallel version of svd using c from scratch..
|
c*******h
发帖数: 1096 | 8 i'll check it out. thank you
【在 s****y 的大作中提到】
: they have more memory。 how about that？
|
k******n
发帖数: 35 | 9 I would try anasazi from Trilinos package. http://trilinos.sandia.gov/packages/anasazi/ |
d******e
发帖数: 7844 | 10 是否试过MATLAB 64位版
试过的话当我没说
【在 c*******h 的大作中提到】
: any (including parallel computing) software can compute
: partail svd on a matrix of size 200,000 x 600,000?
: 300 largest singular values are enough
|
c*******h
发帖数: 1096 | 11 seems that the package is designed for eigen problem and does not
have convenient interface for svd.
for a retangular matrix A, i have to either solve the eigen problem
of AA' or [0 A'; A 0]. The former way requires explictly forming the
matrix AA', which destroys the sparsity. The latter way results in
much larger arnoldi vectors and probably many more iterations.
please correct me if i misunderstood the usage of this package.
【在 k******n 的大作中提到】
: I would try anasazi from Trilinos package. http://trilinos.sandia.gov/packages/anasazi/
|
c*******h
发帖数: 1096 | 12 thanks. see 5478.
i ended up using a machine with 16G mem to solve the problem.
【在 d******e 的大作中提到】
: 是否试过MATLAB 64位版
: 试过的话当我没说
|
d******e
发帖数: 7844 | 13 很好很强大。
【在 c*******h 的大作中提到】
: thanks. see 5478.
: i ended up using a machine with 16G mem to solve the problem.
|
