Parallelisation

With large datasets, retrieval can sometimes take some time. To address this, PyTerrier Transformers can parallelised.

Each Transformer has a .parallel() method, which parallelises the transformer. Two backends are supported:

  • ‘joblib’ - uses multiple processes on your current machine. Resources such as indices will be opened multiple times on your machine. Joblib is the default backend for parallelisation in PyTerrier.

  • ‘ray’ - uses multiple processes on your machine or on other machines in the same cluster, orchestrated in a Ray cluster. Large indices will be reopened on each machine.

Parallelisation occurs by partitioning dataframes and separating them across different processes. Partitioning depends on the type of the input dataframe:

  • queries: partitioned by qid

  • documents: partitioned by docno

  • ranked documents: partitioned by qid

NB: Parallelisation is an experimental features. Please let us know what works or what doesnt work using the PyTerrier issue tracker.

NBB: Parallelisation is known not to work on Windows, and hence is disabled.

Parallelisation using Joblib

A transformer pipeline can be parallelised by using the .parallel() transformer method:

dph = pt.BatchRetrieve(index, wmodel='DPH')
dph_fast = dph.parallel(2)

In this way, any set of queries passed to dph_fast will be separated into two partitions, based on qid, and executed on dph.

Parallelisation using Ray

Ray is a framework for distributing Python tasks across multiple machines. For using it in PyTerrier, setup your Ray cluster by following the Ray documentation. Thereafter parallelisation over Ray can be used in PyTerrier in a similar way as for joblib:

import ray
ray.init() #configure Ray as per your cluster setup
dph = pt.BatchRetrieve(index, wmodel='DPH')
dph_fast = dph.parallel(2, backend='ray')

In particular, ray.init() must have been called before calling .parallel().

What to Parallelise

Only transformers that can be pickled. Transformers that use native code may not be possible to pickle. Some standard PyTerrier transformers have additional support for parallelisation:

  • Terrier retrieval: pt.BatchRetrieve(), pt.FeaturesBatchRetrieve()

  • Anserini retrieval: pt.AnseriniBatchRetrieve()

Pure python transformers, such as pt.text.sliding() are picklable. However, parallelising only pt.text.sliding() may not produce efficiency gains, due to the overheads of shuffling data back and forward.

Entire transformer pipelines (i.e. combined using operators) can be pickled and parallelised. In general, you should parallelise the most inefficient component of your process, while also minimising the amount of data being transferred between processes. For instance, consider the following pipeline:

pipe = pt.BatchRetrieve(index, metadata=["docno", "text"] >> pt.text.sliding() >> pt.text.scorer() >> pt.text.max_passage()

While BatchRetrieve might represent the slowest component of the pipeline, it might make sense to parallelise pipe as a whole, rather than just BatchRetrieve, as then only the queries and final results need to be passed betwene processes. Indeed among the following semantically equivalent pipelines, we expect parallel_pipe0 and parallel_pipe2 to be faster than parallel_pipe1:

parallel_pipe0 = pt.BatchRetrieve(index, metadata=["docno", "text"]).parallel() >> pt.text.sliding() >> pt.text.scorer() >> pt.text.max_passage()
parallel_pipe1 = ( pt.BatchRetrieve(index, metadata=["docno", "text"]).parallel() >> pt.text.sliding() ).parallel(2)  >> pt.text.max_passage()
parallel_pipe2 = pipe.parallel(2)

There are of course overheads on paralllelisation - for instance, the Terrier index has to be loaded for each separate process, so your machine(s) require enough memory. Shared resources such as GPU cards will need careful consideration - adding multiple processes accesssing the same resources will not increase speed and may add problems instead.

Finally, we do not recommend parallelisation in resource-constrained containerised environments such as Google Colab.

If you find PyTerrier transformers that do not parallelise and you think it should, please raise an issue on the PyTerrier github repository.

Outlook

We expect to integate parallelisation at different parts of the PyTerrier platform, such as for conducting a gridsearch. Moreover, we hope that proper integration of multi-threaded retrieval in pt.BatchRetrieve() (while requires upstream improvements in the underlying Terrier platform) will reduce the need for this form of parallelisation.