[Boost-users] RE: Boost thread problem

From: C. Michailidis (dinom_at_[hidden])
Date: 2005-04-14 18:02:07

> It seems the usage of the vector cases problems with the multithreaded 
programs on the multiprocessor machines. 
> I have changed it to queue and the bug has been resolved. I cannot reproduce 
it now.

Huh, that's interesting!  I suppose it may no longer be a major issue for you 
then (even though vectors and queues have different algorithms under the 
hood).

Because you mentioned that the problem seems to be limited to true SMP 
systems, I tried your code again (this time on a dual cpu system) and I still 
couldn't reproduce it.  I must admit that the process didn't seem to bounce 
between cpus though, so perhaps something is lost in the shuffle in that 
regard.  The dual cpu system was configured similarly to the first machine 
that I tried it on (FreeBSD 5.3 + boost 1.32.0_2).  The first machine 
(although not a 'true' smp system) is using hyperthreading w/ an SMP kernel 
so as far as the OS is concerned it too is a multiprocessor machine.

If you still want to isolate the problem that you *were* having (using vector) 
then you may be interested in the details of thread-safe CRTs.

The "-lc_r" indicates that (on FreeBSD) the re-entrant C library (i.e. 
thread-safe) should be linked against.  On Windows I believe the analog is 
called something like 'msvcrtmt.dll' (C RunTime - MultiThreaded).  I would 
hope that the .NET compiler uses this 'msvcrtmt.dll' when 'multithreaded 
support is on', but I couldn't tell you for sure.  If you're statically 
linking, I think the name is just 'msvcrtmt.lib'.  The thread-unsafe version 
doesn't have 'mt' appended to the library filename.

The glamour of MSFT development has been largely replaced with a desire to 
actually get stuff working so... most of my development has been on FreeBSD 
lately.  Therefore some of this MSFT specific information may not be exactly 
correct.  Oh how the cobwebs accumulate!

Anyway, you could probably double check for the 'msvcrtmt' library by looking 
at the detailed linker settings, that is if my suspicion concerns you and 
you're still interested in isolating the problem.

In any case, I'm glad you found a work-around!

C. Michailidis
Balston Research LLC