shared_dict: memory consumption

[ldv1]

Hi,
It is about the piece of code shared_dict.
Reproduced here with minor changes to track the memory consumption:

import resource
from multiprocessing import Manager

import torch
from torch.utils.data import Dataset, DataLoader

def print_mem():
    m = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss/1024
    print("Peak memory (MB):", m)

class MyDataset(Dataset):
    def __init__(self, shared_dict, length):
        self.shared_dict = shared_dict
        self.length = length

    def __getitem__(self, index):
        if index not in self.shared_dict:
            print('Adding data')
            self.shared_dict[index] = torch.tensor([float(index)]*100000)
        return self.shared_dict[index]

    def __len__(self):
        return self.length

# Init
manager = Manager()
shared_dict = manager.dict()
dataset = MyDataset(shared_dict, length=100)

loader = DataLoader(
    dataset,
    batch_size=10,
    num_workers=0,
    shuffle=False,
    pin_memory=False
)

# First loop will add data to the shared_dict
print_mem()
for x in loader:
    pass # print(x)

# The second loop will just get the data
print_mem()
for x in loader:
    pass # print(x)

# The third loop will just get the data too
print_mem()
for x in loader:
    pass # print(x)

print_mem()

To make things easier to understand, the number of workers is set to 0.

If I run this code, the memory keeps increasing after each loop.

If I use the standard dict() instead of manager.dict(), memory does not increase after the first loop, as expected.

Would you know why manager.dict() leads to a memory increase?

Thanks for the help.