#1657 ParallelIndexSets are broken when using bigunsignedint as index type
Metadata
Property | Value |
---|---|
Reported by | Steffen Müthing (steffen.muething@iwr.uni-heidelberg.de) |
Reported at | May 27, 2015 10:03 |
Type | Bug Report |
Version | Git (pre2.4) [cmake] |
Operating System | Unspecified / All |
Last edited by | Dominic Kempf (dominic.r.kempf@gmail.com) |
Last edited at | Jun 30, 2015 12:51 |
Closed by | Dominic Kempf (dominic.r.kempf@gmail.com) |
Closed at | Jun 30, 2015 12:51 |
Closed in version | 2.4 |
Resolution | Fixed |
Comment | by Ansgar in f35437e92d8a4ed0fd8. Thank you! |
Description
I just noticed that parallel PDELab programs fail when using the ISTL AMG solver backend, complaining about not being able to construct a bigunsignedint with negative value (which is due to Ansgar's recent fixes to bigunsignedint and is very reasonable). After some digging around, I think the problem is in line 1320 of remoteindices.hh (https://github.com/dune-project/dune-common/blob/master/dune/common/parallel/remoteindices.hh#L1320), where an index is default-initialized with -1 (probably to distinguish between valid and invalid indices).
I'm pretty lost in that code, however. Could someone with a better knowledge of that code base please take a look?
You can reproduce the problem by compiling dnaplfv from dune-pdelab-howto and running the program with more than one MPI process in parallel.