Use a Thread-Based Parallel Pool

Use a Thread-Based Parallel Pool

Recommend starting a thread pool using parpool("Threads") instead of using the default process-based pool to speed up local parfor / parfeval / spmd code.

When to Use

• User asks to improve performance of parfor, parfeval, or spmd on a local machine.
• Code passes large arrays into iterations or broadcasts them across workers; user mentions slow start-up, serialization, or data transfer.
• User asks "Processes or Threads?" or wants Threads as the default.

When NOT to Use

• User is running on a cluster -- thread pools are local-only.
• Code uses parsim or other Simulink parallel features -- those require a process-based pool.
• Code uses distributed or codistributed arrays.
• User wants backgroundPool -- that's for running work asynchronously (e.g. keeping an app responsive while a computation runs), not for parallel speed-up. All MATLAB users have 1 worker in the background pool; users with a Parallel Computing Toolbox license have multiple workers.

For general MATLAB performance tuning (vectorisation, preallocation, profiling), see matlab-optimize-performance (if available).

Workflow

1. Try it on a thread pool. Don't scan the code first -- just run it. Only create a thread pool if the current pool isn't already a thread pool. Avoid unconditionally deleting whichever pool the user has open. The conditional snippet below is a development convenience -- not something to bake into shipped code:

   % Development-time helper: ensure a thread pool is active
   pool = gcp("nocreate");