[Boost-MPI][Boost-ODEINT] Running odeint within MPI-context without using mpi_state

[Boost - Users mailing list]

Hei,

I am currently using the odeint-solver for calculating the right hand
side of a non-linear equation. For speeding up the calculation of the
right hand side, I decided to test the MPI-implementation. Now, after I
would prefer to keep the code backwards compatible, I did not want to
replace state_type with mpi_state, and instead split up the initial
vector by hand before calling odeint without telling odeint that I am
operating within an MPI context. I take care of all communications which
involve the vectors, and therefore I assumed that this approach should
work (especially after it worked correctly in small test programs).

Nevertheless, I encountered the issue that after a certain amount of
steps the step size between the different threads starts to differ by a
small amount, even though I am in the same iteration. As example:

step number   rank    step size

n                    1        0.00025
n                    0        0.00025 //Ok
n + 1              1        0.00051152
n + 1              0        0.000511523 //Not ok

I checked my program with valgrind for possible memory leaks/corruption
which could overwrite something, but nothing relevant came up. The
problem is repeatable.

Therefore, is the general idea correct at all, or am I doing something
wrong here by neglecting a possible communication between the different
threads within odeint?

Thanks!

Regards,

Roland



_______________________________________________
Boost-users mailing list
[hidden email]
https://lists.boost.org/mailman/listinfo.cgi/boost-users

[Boost - Users mailing list]

Hello
> Therefore, is the general idea correct at all, or am I doing something
> wrong here by neglecting a possible communication between the different
> threads within odeint?

You probably must only use steppers where you decide the step size and
which don't adjust their step size dynamically. Otherwise I guess you'd
have to modify the dynamic step size code to synchronize step sizes
between different mpi processes, etc.
Ilja
_______________________________________________
Boost-users mailing list
[hidden email]
https://lists.boost.org/mailman/listinfo.cgi/boost-users

[Boost - Users mailing list]

So, there is no communication between the threads about the step size,
and each thread can choose their own step size?

Thanks,

regards,

Roland

Am 19.01.21 um 09:52 schrieb Ilja Honkonen:

> Hello
>> Therefore, is the general idea correct at all, or am I doing something
>> wrong here by neglecting a possible communication between the different
>> threads within odeint?
>
> You probably must only use steppers where you decide the step size and
> which don't adjust their step size dynamically. Otherwise I guess
> you'd have to modify the dynamic step size code to synchronize step
> sizes between different mpi processes, etc.
> Ilja

_______________________________________________
Boost-users mailing list
[hidden email]
https://lists.boost.org/mailman/listinfo.cgi/boost-users