|
Документ взят из кэша поисковой машины. Адрес
оригинального документа
: http://www.atnf.csiro.au/computing/software/askapsoft/sdp/docs/current/platform/processing.html
Дата изменения: Unknown Дата индексирования: Tue Apr 12 13:03:25 2016 Кодировка: ISO8859-5 Поисковые слова: п п п п п п п п п п п п п п п п п р п р п п р п п р п |
The ASKAP Central Processor platform consists of a Cray XC30 supercomputer and a 1.4PB Cray Sonexion storage system mounted as a Lustre filesystem.
To login to the ASKAP Central Processor:
ssh abc123@galaxy.ivec.org
Where abc123 is the login name iVec gave you. In that case that your CSIRO login differers from your iVec login, and to avoid having to specify the username each time, you can add the following to your ~/.ssh/config file:
Host *.ivec.org
User abc123
Add the following to your ~/.bashrc
# Use the ASKAP environment modules collection
module use /group/askap/modulefiles
# Load the ASKAPsoft module
module load askapsoft
# Load the ASKAP pipeline scripts module
module load askappipeline
# Load the measures data
module load askapdata
# Load the BBCP module for fast external data transfer
module load bbcp
# Load the "ashell" module for access to the commissioning archive
module load ashell
# Allow MPICH to fallback to 4k pages if large pages cannot be allocated
export MPICH_GNI_MALLOC_FALLBACK=enabled
You have two filesystems available to you:
The scratch filesystem should be used for executing your processing jobs. Specifically bulk data (measurement sets, images, etc) to be read from the job, and all output written, should be from/to the scratch filesystem.
Note that your home directory, while it can be read from the compute nodes, cannot be written to from the compute nodes. It is mounted read-only on the compute nodes to prevent users from being able to “clobber” the home directory server with thousands of concurrent I/Os from compute nodes.
This section describes the job execution environment on the ASKAP Central Processor. The system uses SLURM for Job scheduling, however the below examples use a Cray specific customisation to declare the resources required. An example slurm file is:
#!/usr/bin/env bash
#SBATCH --time=01:00:00
#SBATCH --ntasks=80
#SBATCH --ntasks-per-node=20
#SBATCH --job-name=myjobname
#SBATCH --no-requeue
#SBATCH --export=NONE
aprun -B ./myprogram
Note the use of aprun instead of mpirun. The -B option to aprun tells ALPS (the Application Level Placement Scheduler) to reuse the width, depth, nppn and memory requests specified with the corresponding batch reservation.
Specifically, the following part of the above file requests 80 processing elements (PE) to be created. A PE is just a process. The parameter ntasks-per-node says to execute 20 PEs per node, so this job will require 4 nodes (80/20=4):
#SBATCH --ntasks=80
#SBATCH --ntasks-per-node=20
Then to submit the job:
sbatch myjob.slurm
It may often be the case that you will want to submit a job that depends on another job for valid input (for instance, you want to calibrate a measurement set that is being split from a larger measurement set via mssplit).
The sbatch command allows the specification of dependencies, which act as prior conditions for the job you are submitting to actually run in the queue. The syntax is:
sbatch -d afterok:1234 myjob.slurm
The “-d” flag indicates a dependency, and the afterok: option indicates that the job being submitted (myjob.qsub) will only be run after job with ID 1234 completes successfully. There are other options available - see the man page for sbatch for details.
The ID of a job is available from running squeue. If you are running a script that involves submitting a string of inter-dependent programs, you may want to capture the ID string from sbatch’s output. When you run sbatch, you get something like this:
> sbatch myjob.slurm
Submitted batch job 1234
which you could parse using something like the following (this would run in a bash script - adapt accordingly for your scripting language of choice):
JOB_ID=`sbatch myjob.slurm | awk '{print $4}'`
And you would then use that environment variable in the dependency option:
sbatch -d afterok:${JOB_ID} myjob.slurm
The following example launches a job with a number of PEs that is not a multiple of ntasks-per-node, in this case 22 PEs:
#!/usr/bin/env bash
#SBATCH --time=01:00:00
#SBATCH --ntasks=22
#SBATCH --ntasks-per-node=20
#SBATCH --job-name=myjobname
#SBATCH --no-requeue
#SBATCH --export=NONE
aprun -n 22 -N 20 ./myprogram
Note that instead of passing “-B”, which says use the numbers from ntasks & ntasks-per-node, you must pass “-n” and “-N” specifically. Using the “-B” option only works if ntasks is divisible by ntasks-per-node.
OpenMP Programs:
The following example launches a job with 20 OpenMP threads per process (although there is only one process). The cpus-per-task option declares the number of threads to be allocated per process. The below example starts a single PE with 20 threads:
#!/usr/bin/env bash
#SBATCH --time=00:30:00
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=20
#SBATCH --job-name=myjobname
#SBATCH --export=NONE
# Instructs OpenMP to use 20 threads
export OMP_NUM_THREADS=20
aprun -B ./my_openmp_program
To see your incomplete jobs:
squeue -u $USER
Sometimes it is useful to see the entire queue, particularly when your job is queued and you wish to see how busy the system is. The following commands show running jobs:
squeue
apstat
And to display accounting information, that includes completed jobs, the following command can be used:
sacct
If you wish to cancel a job that is running, or still in the queue, you use the scancel command together with the job ID:
scancel 1234
Any jobs that depend on this one (see above) should also get cancelled at the same time.