In many functions from scikit-learn implemented user-friendly parallelization. For example in
sklearn.cross_validation.cross_val_score you just pass desired number of computational jobs in n_jobs argument. And for PC with multi-core processor it will work very nice. But if I want use such option in high performance cluster (with installed OpenMPI package and using SLURM for resource management) ? As I know sklearn uses joblib for parallelization, which uses multiprocessing. And, as I know (from this, for example, Python multiprocessing within mpi) Python programs parallelized with multiprocessing easy to scale oh whole MPI architecture with mpirun utility. Can I spread computation of sklearn functions on several computational nodes just using mpirun and n_jobs argument?
-
1You might want to check dask-sklearn with the distributed scheduler, that can run in a cluster: jcrist.github.io/dask-sklearn-part-1.htmldukebody– dukebody2016-08-07 10:28:11 +00:00Commented Aug 7, 2016 at 10:28
-
@dukebody can you post an example for using the distributed scheduler? The distributed dask examples I've seen involve manually creating workers on each machine and assigning them to the scheduler. I'm not sure I see how this ties in to the dask-sklearn functions. Would I just create the scheduler and workers like here: dask.pydata.org/en/doc-test-build/distributed.html then set the default scheduler like in your link (where 10.0.0.3:8786 is the address of the scheduler I created like in the first link)?IVlad– IVlad2016-08-07 11:52:37 +00:00Commented Aug 7, 2016 at 11:52
-
Yes. The setup process is exactly as you describe. See distributed.readthedocs.io/en/latest/setup.htmlMRocklin– MRocklin2016-08-07 12:29:09 +00:00Commented Aug 7, 2016 at 12:29
-
@MRocklin that doesn't seem to work for me. It seems that nothing gets executed on the workers, although they are successfully created. Can you read the answer below and my comments to it and see if you have any ideas please?IVlad– IVlad2016-08-08 18:23:18 +00:00Commented Aug 8, 2016 at 18:23
1 Answer
SKLearn manages its parallelism with Joblib. Joblib can swap out the multiprocessing backend for other distributed systems like dask.distributed or IPython Parallel. See this issue on the sklearn github page for details.
Example using Joblib with Dask.distributed
Code taken from the issue page linked above.
from sklearn.externals.joblib import parallel_backend
search = RandomizedSearchCV(model, param_space, cv=10, n_iter=1000, verbose=1)
with parallel_backend('dask', scheduler_host='your_scheduler_host:your_port'):
search.fit(digits.data, digits.target)
This requires that you set up a dask.distributed scheduler and workers on your cluster. General instructions are available here: http://dask.readthedocs.io/en/latest/setup.html
Example using Joblib with ipyparallel
Code taken from the same issue page.
from sklearn.externals.joblib import Parallel, parallel_backend, register_parallel_backend
from ipyparallel import Client
from ipyparallel.joblib import IPythonParallelBackend
digits = load_digits()
c = Client(profile='myprofile')
print(c.ids)
bview = c.load_balanced_view()
# this is taken from the ipyparallel source code
register_parallel_backend('ipyparallel', lambda : IPythonParallelBackend(view=bview))
...
with parallel_backend('ipyparallel'):
search.fit(digits.data, digits.target)
Note: in both the above examples, the n_jobs parameter seems to not matter anymore.
Set up dask.distributed with SLURM
For SLURM the easiest way to do this is probably to use the dask-jobqueue project
>>> from dask_jobqueue import SLURMCluster
>>> cluster = SLURMCluster(project='...', queue='...', ...)
>>> cluster.scale(20)
You could also use dask-mpi or any of several other methods mentioned at Dask's setup documentation
Use dask.distributed directly
Alternatively you can set up a dask.distributed or IPyParallel cluster and then use these interfaces directly to parallelize your SKLearn code. Here is an example video of SKLearn and Joblib developer Olivier Grisel, doing exactly that at PyData Berlin: https://youtu.be/Ll6qWDbRTD0?t=1561
Try Dask-ML
You could also try the Dask-ML package, which has a RandomizedSearchCV object that is API compatible with scikit-learn but computationally implemented on top of Dask
https://github.com/dask/dask-ml
pip install dask-ml
11 Comments
dask-ssh to set up my scheduler and workers. That works fine, if I print the scheduler object I get the right number of cores (240). Next, I wrapped the call to the randomizedsearch's fit in the with statement. If I look in the console window where i executed dask-ssh, I see a connection from the node I run the python script in. However, there is no distributed work going on. It doesn't scale, and it doesn't even see the GPUs that the workers have.n_jobs parameter, setting to -1, 1, 100, 240. Each value above 20 leads to about the same performance, which makes me think that nothing is actually running on the distributed workers, but on the node I run the python script on (gensim also prints a message that there is no GPU. There is a GPU on the worker nodes, but there isn't one on the node I run the script from).
ipyparallel, same thing I described with dask. The workers (engines in ipyparallel) are successfully created, the client sees them, but my grid searches do not run on them.