[Networkit] LAMG not converning for small tolerance

Simon Adrian simon.adrian at tum.de
Wed Nov 9 14:03:17 CET 2016


Dear Henning,

I'm not able to compile the version from the Dev branch (please find my compile 
log attached). 

Thus I wired in the quick fix manually into the master version and as far as I 
can tell, everthing works as it should regarding LAMG.

Thanks a lot for your help.

Best regards,
Simon



Am Montag, 7. November 2016, 09:23:11 CET schrieb Henning Meyerhenke:
> Dear Simon,
> 
> Could you please pull and try the Dev branch? Michael Wegner, the LAMG
> author, has pushed a quick fix. Does it solve your problem?
> 
> Best regards,
> Henning
> 
> Am 06.11.16 um 16:19 schrieb Simon Adrian:
> > Dear Moritz,
> > 
> > thank you for your reply.
> > 
> > The runtime for the entire program is less than 10 seconds, so I would
> > rule
> > this out as the source of the problem.
> > 
> > Nevertheless, I have increased the value by a factor of ten, but as
> > expected, I could not observe any difference.
> > 
> > Best regards,
> > Simon
> > 
> > Am Freitag, 4. November 2016, 19:27:50 CET schrieb Moritz von Looz:
> >> Dear Simon,
> >> 
> >> thanks for your email and example!
> >> 
> >> In both the original unit test and your example code, the maximum
> >> convergence time is set to ten minutes with "status.maxConvergenceTime =
> >> 10
> >> * 60 * 1000". It is to be expected that with increasing precision and
> >> input
> >> size, the time until convergence grows until it no longer fits into the
> >> set
> >> maximum.
> >> 
> >> Do you mean it took unusually long and failed to converge even when
> >> increasing the maximum time?
> >> 
> >> All the best,
> >> Moritz
> >> 
> >> Am 02.11.2016 um 19:04 schrieb Simon Adrian:
> >>> I have played around with the test example
> >>> cpp/numerics/test/LAMGGTest.cpp
> >>> 
> >>>  of the LAMG implementation.
> >>> 
> >>> It works well when the I use the default tolerance (i.e.,
> >>> status.desiredResidualReduction = 1e-6* b.length() /
> >>> (hierarchy.at(0).getLaplacian() * x - b).length();)
> >>> 
> >>> When I decrease the tolerance, it stops working at around 1e-10 for the
> >>> "wing" graph, that is, it does not converge and the output for the final
> >>> residual is "-nan".
> >>> 
> >>> I'm not an expert in the LAMG algorithm, but it seems to me that this is
> >>> not the correct behavior (given that there is some distance to machine
> >>> precision from the tolerance).
> >>> 
> >>> I compiled a slightly modified version without the googletest
> >>> environment
> >>> (see attachment); I presume that this should not effect the observation.
> >>> 
> >>> Thank you very much for your help!
> >>> 
> >>> Best regards,
> >>> Simon
> >>> 
> >>> 
> >>> 
> >>> _______________________________________________
> >>> NetworKit mailing list
> >>> NetworKit at ira.uni-karlsruhe.de
> >>> https://lists.ira.uni-karlsruhe.de/mailman/listinfo/networkit
> >> 
> >> _______________________________________________
> >> NetworKit mailing list
> >> NetworKit at ira.uni-karlsruhe.de
> >> https://lists.ira.uni-karlsruhe.de/mailman/listinfo/networkit
> > 
> > _______________________________________________
> > NetworKit mailing list
> > NetworKit at ira.uni-karlsruhe.de
> > https://lists.ira.uni-karlsruhe.de/mailman/listinfo/networkit

-------------- next part --------------
A non-text attachment was scrubbed...
Name: compile.log
Type: text/x-log
Size: 4126 bytes
Desc: not available
URL: <http://lists.ira.uni-karlsruhe.de/pipermail/networkit/attachments/20161109/c097a61d/attachment.bin>


More information about the NetworKit mailing list