Satish Balay <ba...@mcs.anl.gov> writes: > On Sat, 11 Nov 2017, Matthew Knepley wrote: > >> And that makes my point. The time for me to login to cg, set >> everything up, and run the tests should be automated, and in fact we >> already did that for next, which is what should be used. > > You can automate an ssh command run a test on cg - [if you don't want > to run tests on your laptop] > > But this is a fundamental problem. > > When next model is promoted for integration testing - and Jed keeps > justifying it with all the benefits it would have - and any one of us > short-cirucits it by using it for 'local testing' which I think is a > prereqisite for next model - our next model is broken and will remain > broken. > > Another reason why Barry wants to throw away this model.
It's really just the same reason again. My proposal is to create 1. "pretty good" testing that runs locally in <5 minutes 2. a recommended PETSC_ARCH using an open source compiler (gcc or clang) that finds common bugs; this is probably complex, int64, etc. 3. continuous integration that runs this "pretty good" suite above on weird architectures in a few minutes for all PRs and flags them on the web interface + notifies the author. These are small, desirable practices in any case. If we do this and people are still abusing 'next', we can revisit what has gone wrong.