Skip to content

Do not scan the same files multiple times in ScalableShardDataset #137

Description

@rualark

It seems that ScalableShardDataset can run os.walk of the same folders and then get length of each file potentially hundreds of times depending on the number of logical shards in one rank, because each StreamingDocDataset runs full os.walk() in its setup:

[d.setup() for d in self.data]

for root, dirs, files in os.walk(datapath, topdown=False)

This can be especially inefficient in case of many files. Should os.walk() run only once and its result should be reused?

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions