Shared Memory with Ray

Question

I parallelise a CPU-bound task, which takes a large nested list as read-only input. To avoid that the nested list is repeatedly copied into the processes, I would like to make the object accessible via shared memory.

I am working on a 64-bit Windows machine, using Python 3.8.5 (default, Sep 3 2020, 21:29:08) [MSC v.1916 64 bit (AMD64)] and Spyder 4.2.1.

I already tested implementing shared memory by ray.put(foo) where foo represents a large nested list. However, from the Task manager, I deduce that the nested list is copied into each process.

I was wondering whether Ray offers an alternative solution. In what follows, I present a minimal example to showcase my use-case.

import numpy as np
import ray
import sys 
import time

@ray.remote    
def toy_function(x, nested_list):
    for i in range(5* 10**4):
      m1 = np.random.random((50,50)) 
      matrix = m1 @ np.transpose(m1) 
      sol = np.linalg.inv(matrix)
    result = x + nested_list[0][0]
    return result


if __name__ == "__main__":
    ray.init(num_cpus=5)
    
    # Create a nested list
    N = 10**8
    nested_list = [list(range(2*N)), list(range(3*N))]
    nested_list_id = ray.put(nested_list)
 
    t0 = time.time()
    result = ray.get([toy_function.remote(i, nested_list_id) for i in range(5)])
    t1 = time.time() - t0

    print(f"Execution time: {t1} sec.")
    ray.shutdown()