Re: [Starpu-devel] Failed assert in copy_driver.c

[Xavier Lacoste <xl64100@gmail.com>]

Le 21 oct. 2014 à 16:53, Samuel Thibault <samuel.thibault@ens-lyon.org> a écrit :

Xavier Lacoste, le Tue 21 Oct 2014 15:06:40 +0200, a écrit :

oops in copy_data_1_to_1_generic (datawizard/copy_driver.c:211)... 4: unspecified launch failure

Ok. So since changing the number of requests doesn't fix things, CUDA is
probably not overwhelmed with too many requests.  Is your application
filling GPU memory a lot?  Perhaps there is a race between a data
transfer and a buffer deallocation, is there a way I could run the
program myself to debug things with gdb at ease?

Yes, for that you have to install first scotch (I don't know if I can share my installation, I did a cd $WORKDIR; chmod -R a+rX scotch_5.1.12; chmod -R a+rX ricar with WORKDIR=/ccc/work/cont003/gen1567/lacoste)

my environnement is in /ccc/work/cont003/gen1567/lacoste/ricar/src/loadEnv.cpy

an example of job file:  /ccc/work/cont003/gen1567/lacoste/ricar/src/results_funn_4mpi/curie_hybrid_afshell10_ws+LU+5fact+flop+cmin20+frat8+r11360+bsize60-120_4_8_0_FANIN0__20141020_100031/job.sh

I added you to the Git repository :

git clone git+ssh://thibault@scm.gforge.inria.fr//gitroot/ricar/ricar.git

You need to get the develop branch and you can use my config.in file to configure PaStiX Makefiles.

I can come to help you set this up.

Moreover, 1.2 gives more FLOPS than 1.1 on my application (which is a good evolution :D).

How much more?  Are you using the STARPU_CUDA_ASYNC flag BTW?

One has to read the FLOPS arrays has follow :

First part :  PaStiX without StarPU

Second Part : Fan-in algrithm in StarPU (contribution accumulated into temporary buffers before being sent)

Third part : Fan-out in StarPU (Column Blocks are sent after they are factorized so that the destination column block can be updated by its owner)

Head of columns means MAX_FLOPS_${MPI_NODE_NUMBER}_${GPU_NUMBER} (full node used, i.e. 8 cores or 7 cores + 1 GPUs, or 6+2)

starpu 1.1 (SVN) : https://lut.im/QqbhiacC/b6PV7ZuG

starpu trunk : https://lut.im/vhhtvHbP/IY4ltoe1

Samuel

Attachment: signature.asc
Description: Message signed with OpenPGP using GPGMail