Re: [OMPI users] MPI_Abort under slurm
Hi Ralph, thanks for your answer. I am using: >mpirun --version mpirun (Open MPI) 1.5.4 Report bugs to http://www.open-mpi.org/community/help/ and slurm 2.5. Should I try to upgrade to 1.6.5? /David/Bigagli www.davidbigagli.com On Mon, Feb 25, 2013 at 7:38 PM, Bokassa wrote: > Hi, >I noticed that MPI_Abort() does not abort the tasks if the mpi program > is started using srun. > I call MPI_Abort() from rank 0, this process exit, but the other ranks > keep running or waiting for IO > on the other nodes. The only way to kill the job is to use scancel. > However if I use mpirun under a slurm allocation then MPI_Abort() works as > expected aborting > all tasks. > > Is this a known issue? > > Thanks, David > >
Re: [OMPI users] MPI_Abort under slurm
It should work - check the following srun option: -K, --kill-on-bad-exit[=0|1] Controls whether or not to terminate a job if any task exits with a non-zero exit code. If this option is not specified, the default action will be based upon the SLURM configuration parameter of KillOnBadExit. If this option is specified, it will take precedence over KillOnBadExit. An option argument of zero will not terminate the job. A non-zero argument or no argument will terminate the job. Note: This option takes precedence over the -W, --wait option to terminate the job immediately if a task exits with a non-zero exit code. My guess is that your configuration parameter for KillOnBadExit has not been specified, or you aborted with a zero status. On Feb 26, 2013, at 9:08 AM, Bokassa wrote: > Hi Ralph, thanks for your answer. I am using: > > >mpirun --version > mpirun (Open MPI) 1.5.4 > > Report bugs to http://www.open-mpi.org/community/help/ > > and slurm 2.5. > > Should I try to upgrade to 1.6.5? > > > > /David/Bigagli > www.davidbigagli.com > > > On Mon, Feb 25, 2013 at 7:38 PM, Bokassa wrote: > Hi, >I noticed that MPI_Abort() does not abort the tasks if the mpi program is > started using srun. > I call MPI_Abort() from rank 0, this process exit, but the other ranks keep > running or waiting for IO > on the other nodes. The only way to kill the job is to use scancel. > However if I use mpirun under a slurm allocation then MPI_Abort() works as > expected aborting > all tasks. > > Is this a known issue? > > Thanks, David > > > ___ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users
[OMPI users] Programming with Big Data in R
Hi folks I have someone who is interested in knowing if anyone is currently working with pbdR: http://r-pbd.org/ Specifically, they are trying to evaluate it for their purposes, and would like to hear of any issues and/or other adoption experiences. You can send any comments directly to me, or to the mailing list. Thanks Ralph
Re: [OMPI users] Programming with Big Data in R
On Feb 26, 2013, at 12:17 PM, Ralph Castain wrote: > I have someone who is interested in knowing if anyone is currently working > with pbdR: > It looks to me like an evolution of the capabilities in the `snow' wrapper of `Rmpi', but the addition of the BLACS/PBLAS/ScaLAPACK interfaces data structure accessors. I've used the former quite a bit, but not pbdR itself. Take a look at http://cran.r-project.org/web/views/HighPerformanceComputing.html to get a sense of the kind of packages that are available; there's a lot of overlap, unfortunately. R itself is not a compiled language, but it incorporates routines, standard libraries, and third party packages that package-up C, C++, and Fortran behind the scenes. To the extent one can find a `worker' that ends-up being a mostly native code implementation and runs for a long time, MPI or socket messaging can be useful. Scalars are just length 1 vectors in R, so there's at least the possibility of getting performance by being highly vectorized. pbdR and the others usually provide an `apply' routine that maps a function over a vector. Performance-wise think Python or Perl speed. In contrast to the MPI or sockets, there's a standard package in the distribution called `parallel' that does `fork' of the R process on multicore machines. This works surprisingly well, and if you have a fat node (e.g. 48 processors), it would be my first choice. It's easier to use. Marcus
Re: [OMPI users] Programming with Big Data in R
Also, with regard to your subject line, there are a wide variety of options for connecting to data. Everything from `redis' (http://redis.io/, http://cran.r-project.org/web/packages/rredis/index.html) to HDF5 (http://cran.r-project.org/web/packages/hdf5/index.html), to memory mapped files (http://cran.r-project.org/web/packages/mmap/index.html), to RDBMSs (http://cran.r-project.org/web/packages/RODBC/index.html). With regard to the latter, just slam down some $$$ for a http://www.oracle.com/us/products/database/exadata/overview/index.html and you're good to go. (Wondering if one of these days DB2 or Oracle appliance will be the next big thing in parallel filesystems.) Marcus
