''

OMPR - Object-based Multi-Processing Runner

OMPR is a simple tool for processing tasks with object-based subprocesses.

OMPR may be used for parallel processing of any type of tasks. Usually, a task means a function with a given set of parameters. This function needs to be run (processed) to get a return value - result. There are also scenarios when for processing a given task a (big) object is needed. The problem arises when the time taken by object __init__ is much higher than the time taken by pure processing. An example of such a task is sentence parsing using a SpaCy model. OMPR allows initializing such an object once in each subprocess while forking.

---

Setup

To run OMPR, you will need to:

• Define a class that inherits from RunningWorker. An object of that class will be built in each subprocess. RunningWorker must implement the process(**kwargs) method that is responsible for processing a given task and returning its result. Task parameters and arguments are given with kwargs (dict), and the result may be of any type.
• Build OMPRunner, giving during initialization:
  • RunningWorker type
  • Devices (GPU / CPU) to use
  • Optionally define some advanced parameters of OMPRunner
• Give to OMPRunner tasks as a list of dicts with the process() method. You may give any number of tasks at any time. This method is non-blocking. It just gets the tasks and sends them for processing immediately.

OMPRunner processes given tasks with InternalProcessor (IP) that guarantees a non-blocking interface of OMPRunner. Results may be received with two get methods (single or all) and by default will be ordered according to the tasks' order. Finally, OMPRunner needs to be closed with exit().

This package also delivers the simple_process() function for simple tasks processing when an object is not needed. You can check /examples for sample run code.

If you have any questions or need any support, please contact me: [email protected]

---

More about RunningWorker

There are two policies (managed by OMPR, controlled with the rww_lifetime parameter) of RunningWorker lifecycle:

1st - RunningWorker is closed after processing some task (1..N)
2nd - RunningWorker is closed only when it crashes or with the OMP exit

Each policy has job-specific pros and cons. By default, the second is activated with rww_lifetime=None.

+ all RunningWorkers are initialized once while OMP inits - it saves time
- memory kept by the RunningWorker may grow over time (while processing many tasks)