Multiplying sparse matrices is slow

• Key: MATH-248
• Type: Improvement
• Status: Closed
• Resolution: Fixed
• Priority: Minor
• Assignee: Luc Maisonobe
• Reporter: Christian Semrau
• Fix Version/s: 2.0

Description

The multiplication of sparse real matrices is very slow compared to real matrices: Ten times as slow for size 200, four times as slow for size 400. The time is independent of the number of nonzero entries, because the general algorithm inherited from AbstractRealMatrix is used. I suggest using a specialized multiplication algorithm for matrices that are "sparse enough", walking only over the nonzero entries in one of the matrices.

Comments

[Christian Semrau]

According to my little tests, walking in optimized order and adding the pairwise products to the result matrix is much faster for very sparse matrices, but for a full matrix (every entry set to 1), it is about 50% slower than the current implementation. So the implementation might switch between the two algorithms. Also one might wish to walk the other matrix if it is more sparse than this.

[Luc Maisonobe]

You are right. It is probably even worse when compared to DenseRealMatrix which should be more cache-friendly than RealMatrixImpl.

It is also possible to walk through non-zero elements of both matrices.

Could you provide a patch for this ?

fixed in subversion repository as of r762117
thanks for the report