[CIG-SHORT] Updated PyLith development priorities
Brad Aagaard
baagaard at usgs.gov
Thu May 22 08:26:29 PDT 2008
On Thursday 22 May 2008, Eric Andreas Hetland wrote:
> 1) in the flowchart, "improved PC for kinematic fault condition" leads to
> wave propagation, but in the description it seems that it is needed for
> fault friction models.
The kinematic fault implementation with cohesive cells uses Lagrange
multipliers which results in a saddle point problem. Traditional wave
propagation implementations use a lumped (i.e., diagonal) mass matrix. The
combination of a diagonal matrix and Lagrange multipliers requires special
considerations for efficiently solving the linear system at each time step
(i.e., we need a special preconditioner for the system). Dynamic (fault
friction) with cohesive cells will not use Lagrange multipliers, so we
shouldn't end up with a saddle point problem.
> 2) restart files is indicated as "Difficult" & "computer science (new
> feature)", while adaptive time stepping is "Easy" & "Geophysics (new
> feature)". This seems backward to me, for instance, I know Geofest has both
> features, but the adaptive time stepping is not really that robust. Are
> there standard algorithms for adaptive time stepping (including non-linear
> rheologies) out there? On the other hand, the restart files work extremely
> well in geofest (also commercial codes I have used), so I thought this was
> a solved problem. When I was at MIT, there was a rule against running any
> program on the cluster that did not support checkpointing.
The green circle, blue square, and black diamond describe the level of effort
required. In some cases this is the level of difficulty but in other cases
(e.g., restart files) it is in the amount of work involved.
Checkpointing is absolutely necessary for very long jobs or jobs on large
clusters. It is less important for jobs that run for less than about 8 hrs
and clusters with less than ~100 cores.
I hope these explanations help clarify the listed features and the color
scales.
Brad
More information about the CIG-SHORT
mailing list