AssertErrors for several users -- norm(G * v - curr) < 1e-6 check failing

vlandau commented 3 years ago

I've had at least three Omniscape users confront errors related to the @assert norm(G * v - curr) < <some tolerance> check after Circuitscape gets a solution. There are currently two open issues in Omniscape.jl, here, and here. Something must be going on with the solvers that Circuitscape is using. The error occurs regardless of whether the user specifies CG+AMG or CHOLMOD as the solver.

A few things are still needed:

Determine if Omniscape is indeed using the CHOLMOD server (there have been reports of that solver not having improved speed over CG+AMG, see issue 101 for Omniscape.jl). In theory because this is a direct solver, shouldn't the matrix norm be 0 (unless it's imperfect due to precision limitations)?
Reproduce the error -- a lot of the problems are occurring for very large problems, but presumably only a few actual moving window iterations in Omniscape. I can try to work on reproducing the error for a problem of manageable size

I wanted to open this issue despite the above not being resolved just to get things in motion. Hopefully we can figure this out!

vlandau commented 3 years ago

Omniscape is indeed correctly using the CHOLMOD solver. Solve times were much faster.

ranjanan commented 3 years ago

@vlandau I just need the files so I can reproduce this issue locally and debug. I have commented on both those threads.

vlandau commented 3 years ago

Sounds good -- a colleague of mine ran into the same issue as well, so I can also try to track down his inputs. He was out this week but back Monday.

ranjanan commented 3 years ago

Closing this because both related Omniscape issues are now closed.

Circuitscape / Circuitscape.jl

AssertErrors for several users -- norm(G * v - curr) < 1e-6 check failing #305