This past week, I had a chance to visit some of the scikit-learn developers at Inria in Paris. It was a fun and productive week, and I’m thankful to them for hosting me and Anaconda for sending me there. This article will talk about some changes we made to improve training scikit-learn models using a cluster.

Scikit-learn uses joblib for simple parallelism in many places. Anywhere you pass an n_jobs keyword, scikit-learn is internally calling joblib.Parallel(...), and doing a batch of work in parallel. The estimator may have an embarrassingly parallel step internally (fitting each of the trees in a RandomForest for example). Or your meta-estimator like GridSearchCV may try out many combinations of hyper-parameters in parallel.

You can think of joblib as a broker between the user and the algorithm author. The user comes along and says, “I have n_jobscores, please use them!” Scikit-Learn says, “I have all these embarrassingly parallel tasks to be run as part of fitting this estimator.” Joblib accepts the cores from the user and the tasks from scikit-learn, runs the tasks on the cores, and hands the completed tasks back to scikit-learn.

Joblib offers a few “backends” for how to do your parallelism, but they all boil down to using many processes versus using many cores.

Parallelism in Python

A quick digression on single-machine parallelism in Python. We can’t say up front that using threads is always better or worse than using processes. Unfortunately the relative performance depends on the specific workload. But we do have some general heuristics that come down to serialization overhead and Python’s Global Interpreter Lock (GIL).

The GIL is part of CPython, the C program that interprets and runs your Python program. It limits your Python process so that only one thread is executing Python at once, defeating your parallelism. Fortunately, much of the numerical Python stack is written in C, Cython, C++, Fortran, or Numba, and may be able to release the GIL. This means your “Python” program, which is calling into Cython or C via NumPy or pandas, can get real thread-based parallelism without being limited by the GIL. The main caveat here is that manipulating strings or Python objects (lists, dicts, sets, etc) typically requires holding the GIL.

So, if we have the option of choosing threads or processes, which do we want? For most numeric / scientific workloads, threads are better than processes because of shared memory. Each thread in a thread-pool can view (and modify!) the samelarge NumPy array or pandas dataframe. With multiple processes, data must be serialized between processes (perhaps using pickle). For large arrays or dataframes this can be slow, and it may blow up your memory if the data is a decent fraction of your machine’s RAM. You’ll have a full copy in each processes.

See Matthew Rocklin’s article and David Beazley’s page if you want to learn more.

Distributed Training with dask.distributed

For a while now, you’ve been able to use dask.distributed as a backend for joblib. This means that in most places scikit-learn offers an n_jobs keyword, you’re able to do the parallel computation on your cluster.

This is great when:

  1. your dataset is not too large (since the data must be sent to each worker);
  2. the runtime of each task is long enough that the overhead of serializing the data across the network to the worker doesn’t dominate the runtime; and
  3. you have many parallel tasks to run (else, you’d just use a local thread or process pool and avoid the network delay).

A good example of this may be fitting a RandomForest. Each tree in a forest may be built independently of every other tree. This next code chunk shows how you would fit a RandomForest using a cluster, though as discussed later this won’t work on the currently released versions of scikit-learn and joblib.


The .fit call is parallelized across all the workers in your cluster. Here’s the distributed dashboard during that training.

 

The center pane shows the task stream as they complete. Each rectangle is a single task, building a single tree in a random forest in this case. Workers are represented vertically. I had 8 workers with 4 cores each, which means up to 32 tasks can be processed simultaneously. We fit the 200 trees in about 20 seconds.

Changes to Joblib

Above, I said that distributed training worked in most places in scikit-learn. Getting it to work everywhere required a bit more work, and was part of last week’s focus.

First, dask.distributed‘s joblib backend didn’t handle nested parallelism well. This may occur if you do something like


Previously, that caused deadlocks. Inside GridSearchCV, there’s a call like

where fit_estimator is a function that itself tries to do things in parallel
So the outer level kicks off a bunch of joblib.Parallel calls, and waits around for the results. For each of those Parallelcalls, the inner level tries to make a bunch of joblib.Parallel calls. When joblib tried to start the inner ones, it would ask the distributed scheduler for a free worker. But all the workers were “busy” waiting around for the outer Parallel calls to finish, which weren’t progressing because there weren’t any free workers! Deadlock!

dask.distributed has a solution for this case (workers secede from the thread pool when they start a long-running Parallelcall, and rejoin when they’re done), but we needed a way to negotiate with joblib about when the secede and rejoin should happen. Joblib now has an API for backends to control some setup and teardown around the actual function execution. This work was done in Joblib #538 and dask-distributed #1705.

Second, some places in scikit-learn hard-code the backend they want to use in their Parallel() call, meaning the cluster isn’t used. This may be because the algorithm author knows that one backend performs better than others. For example, RandomForest.fit performs better with threads, since it’s purely numeric and releases the GIL. In this case we would say the Parallel call prefers threads, since you’d get the same result with processes; it would just be slower.

Another reason for hard-coding the backend is if the correctness of the implementation relies on it. For example, RandomForest.predict preallocates the output array and mutates it from many threads (it knows not to mutate the same place from multiple threads). In this case, we’d say the Parallel call requires shared memory, because you’d get an incorrect result using processes.

The solution was to enhance joblib.Parallel to take two new keywords, prefer and require in Joblib #602. If a Parallelcall prefers threads, it’ll use them, unless it’s in a context saying “use this backend instead”, like


On the other hand, if a Parallel requires a specific backend, it’ll get it.

This is an elegant way to negotiate a compromise between:

  1. the user, who knows best about what resources are available, as specified by the joblib.parallel_backend context manager; and
  2. the algorithm author, who knows best about the GIL handling and shared memory requirements.

After the next joblib release, we’ll update scikit-learn to use these options in places where the backend is currently hardcoded. My example above used a branch with those changes.

Look forward for these changes in the upcoming joblib, dask, and scikit-learn releases. As always, let me know if you have any feedback.