Thanks a lot Karl! Well, I don't usually use PETSc runtime options. I have only a subset of parameters (PC types, KSP types, etc.) which are interfaced by my solver and exposed to the end-user. Other few settings (matrix types for instance) are hardcoded in some objects (which my linear system solver interface makes use of).
I can setup a testcase, but it cannot be too simple. You'll have to install my code to run it. I can setup a script for this. If you want to give it a try, I'll send you some instructions tomorrow (here is night already). Best regards Andrea On Jan 21, 2014, at 10:38 PM, Karl Rupp <[email protected]> wrote: > Hi Andrea, > > > sorry to bother you with this again, but can you please give me a final >> word on whether KSPGMRES in combination with PCASM is currently supposed >> or not to work with MATMPIAIJCUSP on multiple-GPUs as it does (that I >> can clearly see) on 1 GPU? Do you have a test for that? > > There are certainly examples around which allow this functionality, but > multi-GPU algorithms are not routinely tested yet. > > >> Again, for what I can see, residuals are inconsistent with 1 or more >> GPUs and you don't get the same discrepancy with, for instance, PCJACOBI >> (same exact residuals on single and multi-GPU). This makes me think >> (maybe wrongly) that my part of the interfacing with PETSc is right. > > Do you happen to have a simple example to reproduce? Which runtime options do > you use/set? I'll try to reproduce this tomorrow. > > Best regards, > Karli >
