I haven't had time to look at the code in depth, but I guess it comes down to what this

_dss->ElementAt(i - 1, j - 1)

part does. Does it work like the code expects it to?

Either way, this is definitely not how a sparse dot product should be implemented; it is far to inefficient.
Normally, you would use the knowledge and internals of the current sparse structure to skip all the zero elements.

Thanks Borek, you're right (we have modified mitc shells). I can provide a model with beams, please see attachments.
One model is with DSS, the other one (ending with B) with default solver. We get very different results by comparing the .out files.

At this stage, we want to know if our implementation of the missing "time(const FloatArray &x, FloatArray &answer)" in DSS could be considered correct, or if we broke something with it.

HI,
I have taken a look on the problem.
First, below the corrected code for your implementation:

void DSSMatrix :: times(const FloatArray &x, FloatArray &answer) const
{
int i, j, dim;

    dim = this->_sm->neq;

    answer.resize(dim);
        answer.zero();

    for (i = 1; i <= dim; i++) {
        for (j = 1; j <= dim; j++) 
        {
            answer.at(i) += _dss->ElementAt(i - 1, j - 1) * x.at(j);  
        }
    }
}

As pointed out by Mikael, this is not really efficient way. However, the dss sparse matrix is assembled directly into its block structure, which is efficient for factorization, but unfortunately not efficient for implementing the multiplication, as the blocks have to be identified (see implementation of ElementAt method) when traversing rows. There is workaround for this in DSS, where the matrix can be assembled in compressed column format first and then converted to block format before factorization, but this essentially requires to keep two versions of the matrix in the memory.

My proposal here is to modify nlineeardynamic problem in a way, that mass matrix will be always assembled using default storage, as the only required operation is the multiplication by a vector. We should allow to change storage only for effectiveStiffnessMatrix which is going to act as LHS of linear system. This will lead to more efficient solution.

If you agree, I can make the modification.

I have committed the change, I finally used compressed column format to store mass matrix, which gives optimal performance in terms of memory and performance.