Substantial slow-down in Tpetra Muelu
Created by: spdomin
Greetings,
Does anyone on the Trilinos development team have a recent set of Muelu/ML comparisons?
As I transition fully from ML to Muelu (this effort is part of Epetra deprecation) almost every single test is O(30)% slower. In most cases, I am using a Muelu XML file that “matches” the previous ML input. See attached test example (nonIsoElemOpenJet). Although I am using the blessed XML for Muelu that matches ML, linear solves increased by almost three iterations. Solver time, however, is almost 2x slower. Moreover, preconditioner setup time is also 2x slower. This result is very much representative of all ten+ tests that I am converting.
I was just wondering if anyone on the Trilinos team is keeping up with the previous Muelu benchmarks against ML. At one point, Muleu was only 10% slower in most cases.
Best,
Stefan