Re: distributed cholesky on systemml

well, SystemML decides the execution type of each operation based on its worst-case memory estimate and the available driver memory budget to avoid unnecessary overheads for distributed operations. Maybe the size of the matrix that is fed into cholesky is smaller than the intermediates used for dat

Re: distributed cholesky on systemml

thanks for the context Jeremy - that helps. I also had an offline conversion with Sasha and he pointed me to a script that does exactly that (iterative invert_lower_triangular) combined with a parfor over independent blocks. We'll merge these scripts soon and I'll reach out individually as necessar

Jenkins build is back to stable : SystemML-DailyTest #1598

See

Re: distributed cholesky on systemml

sure no problem - thanks again for catching this issue that was hidden for a while. Yes, the same depth-first characteristic applies to the Cholesky function as well. In contrast to U_triangular_inv, however, there are data dependencies between the blocks per level (at least in the current algorit