Question

我设置了一个IPython并行ipcluster来使用Sun Grid Engine，看起来效果很好：

ipcluster start -n 100 --profile=sge

2016-07-15 14:47:09.749 [IPClusterStart] Starting ipcluster with [daemon=False]
2016-07-15 14:47:09.751 [IPClusterStart] Creating pid file: /home/USERNAME/.ipython/profile_sge/pid/ipcluster.pid
2016-07-15 14:47:09.751 [IPClusterStart] Starting Controller with SGEControllerLauncher
2016-07-15 14:47:09.789 [IPClusterStart] Job submitted with job id: u'6354583'
2016-07-15 14:47:10.790 [IPClusterStart] Starting 100 Engines with SGEEngineSetLauncher
2016-07-15 14:47:10.826 [IPClusterStart] Job submitted with job id: u'6354584'
2016-07-15 14:47:40.856 [IPClusterStart] Engines appear to have started successfully

然后我使用

从笔记本连接

rc = ipp.Client(profile='sge')

但是当我使用并行魔法

时

%%px
from mpi4py import MPI

comm = MPI.COMM_WORLD
nprocs = comm.Get_size()
rank = comm.Get_rank()

print('I am #{} of {} and run on {}'.format(rank,nprocs,MPI.Get_processor_name()))

我所有进程只返回rank 0：

[stdout:0] I am #0 of 1 and run on compute-8-13.local
[stdout:1] I am #0 of 1 and run on compute-8-13.local
[stdout:2] I am #0 of 1 and run on compute-3-3.local
[stdout:3] I am #0 of 1 and run on compute-3-3.local
[stdout:4] I am #0 of 1 and run on compute-3-3.local
...

以下是我的设置脚本：

ipcluster_config.py：

c.IPClusterEngines.engine_launcher_class = 'SGEEngineSetLauncher'
c.IPClusterStart.controller_launcher_class = 'SGEControllerLauncher'
c.SlurmEngineSetLauncher.batch_template_file = '/home/USERNAME/.ipython/profile_sge/sge.engine.template'
c.SlurmControllerLauncher.batch_template_file = '/home/USERNAME/.ipython/profile_sge/sge.controller.template'

ipcontroller_config.py：
```
c.HubFactory.ip = '*'
```

sge.controller.template

# /bin/sh
#$ -S /bin/sh
#$ -pe orte 1
#$ -q sThC.q
#$ -cwd
#$ -N ipyparallel_controller
#$ -o ipyparallel_controller.log
#$ -e ipyparallel_controller.err
module load gcc/5.3/openmpi 
source activate parallel
ipcontroller --profile-dir={profile_dir}

sge.engine.template

# /bin/sh
#$ -S /bin/sh
#$ -pe orte {n}
#$ -q sThC.q
#$ -cwd
#$ -N ipyparallel_engines
#$ -o ipyparallel_engines.log
#$ -e ipyparallel_engines.err

module load gcc/5.3/openmpi
source activate parallel
mpiexec -n {n} ipengine --profile-dir={profile_dir} --timeout=30

Answer 1

自己找到解决方案/错误：

在ipcluster_config.py中的

，我忘了重命名Slurm - ＆gt; SGE的一些案例，所以它应该是

c.IPClusterEngines.engine_launcher_class = 'SGEEngineSetLauncher'
c.IPClusterStart.controller_launcher_class = 'SGEControllerLauncher'
c.SGEEngineSetLauncher.batch_template_file = '/home/USERNAME/.ipython/profile_sge/sge.engine.template'
c.SGEControllerLauncher.batch_template_file = '/home/USERNAME/.ipython/profile_sge/sge.controller.template'

此导致ipcluster使用某种默认的SGE模板，提交100个单独的作业，而不是100个进程的一个作业。

现在我得到了所希望的：

[stdout:0] I am #5 of 100 and run on compute-5-17.local
[stdout:1] I am #9 of 100 and run on compute-5-17.local
[stdout:2] I am #1 of 100 and run on compute-5-17.local
[stdout:3] I am #7 of 100 and run on compute-5-17.local
[stdout:4] I am #2 of 100 and run on compute-5-17.local
...

Sun Grid Engine上的ipcluster仅排名0

1 个答案: