Ask Your Question
1

NEW VERSION: Sparse matrix stores something for each row?

asked 2018-12-18 19:01:58 -0500

imakhlin gravatar image

updated 2018-12-21 19:39:03 -0500

NEW VERSION. This is what this question has boiled down to upon further investigation. Everything below this paragraph in boldface and the subsequent code is the old version and serves purely historical purposes =) I hope this was the right thing to do rather than creating a new question. The issue is that the amount of memory occupied by a sparse matrix seems to depend linearly on the number of rows in the matrix. In particular, my setup wont even let me create a zero $10^9\times 1$ sparse matrix but has no problem with a zero $1\times 10^9$ sparse matrix. This can all be illustrated by running the below code (and it's hardly the way things are supposed to be).

print get_memory_usage()
B=matrix(QQ, 1, 1000000000, {})
print get_memory_usage()
B=matrix(QQ, 10000000, 1, {})
print get_memory_usage()
B=matrix(QQ, 1000000000, 1, {})
print get_memory_usage()

I'm running the following code. Here I create a sparse $10^5\times 10^5$ identity matrix and then apply it repeatedly to a vector in $\mathbb R^{10^5}$ with 100 nonzero coordinates (which is stored as $10^5\times 1$ sparse matrix). Each time I add the result to a list until I run out of 2 GB of memory.

A=matrix(QQ, 100000, 100000, {(i,i):1. for i in range(100000)})
print get_memory_usage()
B=[matrix(QQ, 100000, 1, {(i,0):1. for i in range(100)})]
while (get_memory_usage()<2000): B.append(A*B[-1])
print len(B)
print get_memory_usage()
del B
del A
print get_memory_usage()

I'm receiving (on a freshly started kernel)

204.54296875
196
2003.51953125
1399.0234375

This raises two questions.

  1. Why is there so much memory (1.4 GB) still in use after I successfully ran the code and deleted both objects I've created? That's a leak, right?

  2. Why does deleting a list of 196 sparse matrices with 100 nonzero elements each free up 600 MB? Each such matrix should only take up a few KB, right?

I'm on Windows 8.1/SAGE 8.4.

UPDATE. Transposing the matrices, i.e. writing

...
B=[matrix(QQ, 1, 100000,  {(0,i):1. for i in range(100)})]
for i in range(200): B.append(B[-1]*A)
...

seems to work well memory-wise, it returns

214.50390625
201
215.53125
186.62109375

However, it takes up much more time than the first version. This is probably due to the implementation of sparse matrix multiplication unfortunately containing a loop over the columns of the right matrix. Is there any simple way around this high memory/low speed dilemma?

UPDATE 2. This might not be a memory leak and have more to do with the implementation of sparse matrices in general rather than their multiplication in particular. Apparently, a sparse matrix stores something for each of its rows as shown by

print get_memory_usage()
B=matrix(QQ, 1, 10000000, {})
print get_memory_usage()
B=matrix(QQ, 10000000, 1, {})
print get_memory_usage()

165.45703125
165.51953125
1082.68359375

This has got to be a known issue. I was not able to find a discussion, however. (This might be what is known as csr_matrix in scipy but why this would be chosen as the general standard here is beyond me.)

edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted
0

answered 2018-12-20 14:39:30 -0500

tmonteil gravatar image

This might be related to this ask question, and is tracked on trac ticket 26532.

edit flag offensive delete link more

Comments

1

@tmonteil Yes, I've seen the question and the ticket and they are somewhat related. This is a separate issue though, a memory efficiency issue rather than time efficiency. Plus it concerns not only multiplication but construction of sparse matrices in general, which I realized after posting and tried to explain in the second update. Maybe I should make that clear at the very top.

imakhlin gravatar imageimakhlin ( 2018-12-21 19:04:34 -0500 )edit
1

I think it is quite common to implement matrices with sparse rows or sparse columns, with the idea that a completely zero row or column probably doesn't need to be stored at all, see e.g. https://docs.scipy.org/doc/scipy/refe... That's just the implementation sage offers right now. Perhaps it will grow other sparse forms if someone feels the need to provide them.

nbruin gravatar imagenbruin ( 2018-12-22 02:58:06 -0500 )edit
1

Sounds like OP/Sage may want to follow this advice from the scipy coo_matrix documentation: "COO is a fast format for constructing sparse matrices. Once a matrix has been constructed, convert to CSR or CSC format for fast arithmetic and matrix vector operations."

rburing gravatar imagerburing ( 2018-12-22 10:12:36 -0500 )edit
1

@nbruin I am certainly no expert on sparse matrix algorithms but I still feel like this is a questionable choice for the default (and only) format.

with apologies of nbruin: It looks like I edited this comment when I was intending to reply. If someone can find the history of this comment and restore it in its old form, that would be great.

imakhlin gravatar imageimakhlin ( 2018-12-23 10:10:20 -0500 )edit
1

Another way to get a sparse data structure mimicking a sparse matrix is to use a bivariate polynomial sum a[i,j]x^iy^j. Addition should work pretty well on that (there might be limitation on the exponent size, though).

nbruin gravatar imagenbruin ( 2018-12-23 13:08:36 -0500 )edit

Your Answer

Please start posting anonymously - your entry will be published after you log in or create a new account.

Add Answer

Question Tools

1 follower

Stats

Asked: 2018-12-18 18:20:27 -0500

Seen: 111 times

Last updated: Dec 21 '18