[petsc-dev] Making coalesced mallocs a runtime option

[Lisandro Dalcin]

I've put together an initial implementation:
https://bitbucket.org/petsc/petsc/branch/dalcinl/malloc-coalesced

On 16 April 2015 at 11:49, Lisandro Dalcin <dalcinl at gmail.com> wrote:
> In on-going discussion about malloc(0), I made a mention about this
> approach.  As this idea is somehow independent of the malloc(0) stuff,
> I'm starting a new thread.
>
> Folks, in the attached file you can see how easy is to support
> coalesced mallocs at runtime. This would enable us to disable the
> coalesced mallocs for optimized builds and run the code under valgrind
> to catch invalid memory accesses.
>
> The code has memory leaks on purpose to get output from -malloc_dump,
> you can try it with debug/opt builds, passing  -malloc_coalesced [0|1]
> option and value. If you pass "-n size1,size2" you can control the
> sizes for the PetscMalloc2 calls.
>
> If you use a optimized build, and any of the sizes is 0, and you pass
> -malloc_coalesced 0, valgrind will be able to catch the invalid
> writes.
>
> $ PETSC_ARCH=arch-linux2-c-opt make TestMalloc
>
> $ valgrind -q ./TestMalloc -n 0,1 -malloc_coalesced 1 2>&1 | grep
> "Invalid write"
> <no-ouput>
>
> $ valgrind -q ./TestMalloc -n 0,1 -malloc_coalesced 0 2>&1 | grep
> "Invalid write"
> ==2304== Invalid write of size 1
> ==2304== Invalid write of size 4
> ==2304== Invalid write of size 8
>
> $ valgrind -q ./TestMalloc -n 0,0 -malloc_coalesced 0 2>&1 | grep
> "Invalid write"
> ==2306== Invalid write of size 1
> ==2306== Invalid write of size 1
> ==2306== Invalid write of size 4
> ==2306== Invalid write of size 4
> ==2306== Invalid write of size 8
> ==2306== Invalid write of size 8
>
>
> IMHO, we should definitely implement this approach. If you agree, let
> discuss about when to enable/disable the coalesced mallocs (e.g., if
> users pass -malloc or -malloc_dump, perhaps we should disable
> coalescing, and only enable it back if -malloc_coalesced 1 is
> explicitly asked for).
>
>
> --
> Lisandro Dalcin
> ============
> Research Scientist
> Computer, Electrical and Mathematical Sciences & Engineering (CEMSE)
> Numerical Porous Media Center (NumPor)
> King Abdullah University of Science and Technology (KAUST)
> http://numpor.kaust.edu.sa/
>
> 4700 King Abdullah University of Science and Technology
> al-Khawarizmi Bldg (Bldg 1), Office # 4332
> Thuwal 23955-6900, Kingdom of Saudi Arabia
> http://www.kaust.edu.sa
>
> Office Phone: +966 12 808-0459



-- 
Lisandro Dalcin
============
Research Scientist
Computer, Electrical and Mathematical Sciences & Engineering (CEMSE)
Numerical Porous Media Center (NumPor)
King Abdullah University of Science and Technology (KAUST)
http://numpor.kaust.edu.sa/

4700 King Abdullah University of Science and Technology
al-Khawarizmi Bldg (Bldg 1), Office # 4332
Thuwal 23955-6900, Kingdom of Saudi Arabia
http://www.kaust.edu.sa

Office Phone: +966 12 808-0459