Re: [Starpu-devel] StarPU changeable sizes and StarPU purge

[Samuel Thibault <samuel.thibault@inria.fr>]

Hello,

Rabab Al-omairy, le dim. 01 déc. 2019 22:03:43 +0300, a ecrit:
> after matrix compression each tile ends up having different sizes.
[...]
> However, on distributed memory on which tiles are 2D cyclically
> distributed among mpi nodes I have  the following error:
> 
> [1][starpu_mpi][_starpu_mpi_test_detached_requests][assert failure] MPI_Test
> returning MPI_ERR_TRUNCATE: message truncated

Mmm, indeed, the distributed support for matrices was not extended to
support tiles with varying size. Here the receiver is expecting a full
size since it doesn't know the size is smaller.

I have added some experimental support in the master. In order to try it
you need to uncomment the //#define DYNAMIC_MATRICES in both
src/datawizard/interfaces/matrix_interface.c and mpi/src/starpu_mpi_datatype.c

That is however for now incompatible with OOC support with O_DIRECT
enabled.

> For some matrix size which I can say matrix with medium size, I
> have the following message which cause application to stuck and not continue
> execution.
> 
> [starpu][starpu_memchunk_tidy] Low memory left on node RAM 0 (6405MiB over
> 128808MiB).

So StarPU is lacking memory. Do you have OOC enabled?

Samuel