Closed gukejun1 closed 11 months ago
The --memory-map
flag minimizes the usage of RAM, in exchange for larger usage of disc space, see issue https://github.com/facebookresearch/dlrm/issues/201.
It relies on Fisher-Yates-Rao shuffle algorithm, with an example of its application to shuffling large data sets here.
How is the memory-map function implemented, how to reduce the memory usage principle, and where is it processed in the dlrm code?