Minutes 10 July 2006
Present: Shane, Karl, Travis.
An "audio recording":ftp://ftp.bioeng.auckland.ac.nz/cmiss/opencmiss/meeting-2006-07-10.ogg is available.
Audio recording of meetings
- This is the last meeting that will be recorded unless more interest is shown.
MPI Test problem proposal and implementation
- cm performs much better with "grid-based fem" than with the original
advection-diffusion fem code.
- The problem solved is not the test problem (initial conditions are not set,
boundary conditions are Neumann, and there is an Euler integration of a
cubic cell model) but provides an indication of what to expect if the test
problem were solved similarly.
- a 1000x1000 element problem with 50 time steps of 0.001 and a maximum of 4 iterations in each CG solve took 82 seconds using 1.1 GB resident memory (2.6 virtual). The single matrix assembly took 19.4 seconds.
- A similar libmesh problem but with matrix assembly at each time step took 379 seconds using 1.3 GB resident memory (1.4 virtual).
- The problem solved is not the test problem (initial conditions are not set,
boundary conditions are Neumann, and there is an Euler integration of a
cubic cell model) but provides an indication of what to expect if the test
problem were solved similarly.
- Sundance
- Doesn't have a high level time-stepping object.
- Nice high level classes for describing the mathematics of the problem (in a variation form).
- Easy to get it to work, even for a problem for which it may not have been intended.
- Results on Using Sundance seem reasonable in comparison to libmesh given that it is doing more iterations in each linear solve.
- Code that uses Sundance can be written in an easy to read manner.
- Currently provides only Triangular/tet elements but quads/hexs shouldn't be a problem.
- Not sure about basis functions in addition to Lagrange.
- No AMR.
- A few seg faults happening
- solver reaches max iterations
- running with 4 processes
- Feels a bit like still an academic project; not sure what applications it is being used for.
- Used for Navier Stokes.
- High level operators provided by epetera
- Perhaps this could be imported to other packages.
Libmesh
- Shane wants to establish a dialog with libmesh developers and suggests
asking about assembling only once.
- Karl says this might be in documentation; we should read this first. If we want to establish a dialog then we should get to the point and explain our position.
- We want to appear as developers (not like people building a Windows gui).
- We could ask what would be involved in parallelization of mesh
What about other packages?
- LifeV and freepooma have not yet been considered.
- Do we check these out now before spending more time on libmesh?
- Have we enough evidence that using an existing package is sensible?
Next meeting 11am Thursday 20 July.