LEMON writer fails about tmlqcd

[deuzeman]

Could be machine specific -- I always get failures on our cluster because the ROM-IO components of the MPI system haven't been compiled properly. Could also be a serious issue in Lemon itself, so I'll have a look.

[kostrzewa]

Hmm, ok. This was with openmpi on my laptop. If it works for you the problem is probably here.

[deuzeman]

I'm not even sure I have everything properly installed myself, so I'll have to test a couple of things...

[Siebren]

Hey there, today David reported the same bug on another machine. I am going to have a look at this now. If you have any other additional information, please let me know.

[kostrzewa]

Sorry, can't tell you anything beyond what it says above. Here's the config.log for the lemon library I was using: https://gist.github.com/2170297

[Siebren]

The problem is caused inside the last commit, checking out the previous version is fine. So the quick fix is now to roll back the last commit, to make sure that the head is in a stable state. Then we can fix up the things that were added in the last commit and re-apply once fixed.

An easy test is the lemon_benchmark application that is also installed. That also fails, so an easy check is something like
mpirun -np 4 lemon_benchmark 4 1

So, I'll put this issue inside Lemon, then ask Albert to roll that commit back. When that is done, the issue here on tmLQCD can be closed I think. But please verify that this works with commit 5306a90a4bc1ed47268e156ea1253e96ee03799a for you.

[kostrzewa]

Ok, will try to get this done ASAP, but I will have to figure out all the parameters of the build that I did back then because I don't have that particular build directory anymore.

[deuzeman]

This was a little careless on my side, but I wasn't aware that this version of lemon was being used! It contains some completely unchecked code -- well, I guess it is somewhat tested now :S. At any rate, I have rolled back all changes having to do with prime factorization in etmc/master. The later commits are in a separate branch unstable, which should reduce the changes of accidents like this.

In the current head revision, there is a known bug related to integer overflow for large file sizes. It's not a trivial bug to fix, because the error is related to a known flaw in the MPI-IO interface. I had been working on a fix -- hence the broken code -- but didn't find the time to wrap it up yet.

[Siebren]

Today I am doing clean installs of head revisions of Lemon and tmLQCD with both Tiago (in Groningen on the BG/P), and David (on Huygens in Amsterdam). We run into several interesting issues because of this. Once we have tracked them down further, we will file them as bugs. I think that this one can be closed now, thanks!