[Boost-users] openmpi with boost::log and some late endl

Subject: [Boost-users] openmpi with boost::log and some late endl
From: Peter Vollmer (vollmepr_at_[hidden])
Date: 2016-05-11 04:16:56

I'm working on some Linux systems and boost:log. More in detail:

After replacing a handmade solution for logging with boost:log and
running it with mpi on a single computer I can observe a strange
behavior in console. 

Normally as i understand it, with autoflush turn on, after streaming a
message into the logger it follows a flush followed by an endl and
flush. 

At my scenario after streaming messages on different mpi processes, it
seem that the system make a thread change right after flushing my
message not after the endl flush. The missing automatic endl isn't lost
it will be printed directly after or later but it will be printed. 
One indication that it has to be a thread change of the system is that
an increase of the usable mpi processes increases the presents
dramatically. For example on 4 core 8 thread system it occurs with 4
mpi processes something around 1 or 2 times, with 8 it occurs more than
20 times. 

Is there a solution to prevent this?

Thanks

Peter