Segfault with dash::Array using CSR-Pattern and odd number of units

[BenjaProg]

Hey @dash-project/developers ,
There is some strange behaviour with my test scenario (main.cpp).
CSRBug.zip

I get a segmentation fault if I execute the test with an odd number of units and at least three.
In the loop lines 42-45 is where the segmentation fault happens.
Unfortunately I and @ddiefenthaler couldn't track it further down until now.
(Note: this fault hasn't occured on ddiefenthaler's machine.)

If I enabled the count variable and(!) the cout in the loop the fault is prevented (at my machine...).
So If there is a segfault, the line DASH_LOG_DEBUG( "Message: ", "made it." );
isn't reached anymore.

The variant_x.txt files hold some of the output possibilites on my machine.
With variant_1 and variant_3 not all processes did return and i had to send a interrupt signal.

Thanks for your help in advance!

[devreal]

Can you please provide us with a stack trace of the Segfault? If you run locally you can do something like:

mpirun -n 3 xterm -e gdb --args ./a.out <parameter>

[BenjaProg]

okay i will provide it this afternoon, unfortunately i am on my way to university every moment

[rkowalewski]

And please provide your build settings: Which compiler, which libraries, etc.

[fuchsto]

So If there is a segfault, the line DASH_LOG_DEBUG( "Message: ", "made it." );
isn't reached anymore.

You don't know, at least not from missing output. The log operation might have been completed but a subsequent segfault prevented the flush of std::clog.

[BenjaProg]

sorry for the delay, I gladly provide you with a stack trace but I first have to work into gdb more.
@rkowalewski: hope the attached info from the build script hold all relevant infos ;-)
additional Information_1.txt

[BenjaProg]

That's the backtrace I got.
I am sorry for the screenshot! Is there a way to pipe the output of gdb in a file?
Because copy/paste wasn't possible.

[devreal]

I just gave it a try with current development and cannot reproduce the issue locally. I used 1.10.2, GCC 5.4.1 and 3 processes. Also tried with 5 processes, still succeeded. Missing something?

[BenjaProg]

@devreal Probably not, it seems the error only appears on my local vm :/
"1.10.2" is the version number of what?
I will have a closer look this weekend.

[devreal]

Ahh sorry, yes. I meant OpenMPI 1.10.2, which is the version that comes packaged with my system.

[fuchsto]

@BenjaProg You could test local sizes before assigning values. Are values in local_sizes[myid] identical to array.lend() - array.lbegin()?

[BenjaProg]

There is now a dedicated branch for this bug. A CSRPatternTest is already there and array.lend() - array.lbegin() is being checked as well, but hasn't been a problem so far.

[fuchsto]

The actual defect is in allocation in DART, see #280 and CSRPatternTest in #279