Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description of changes:
Previously, the range of each shard in the
py1e
shuffle algorithm was not randomized. So when training on a large number of nodes, each node would need to download NCN/PN number of shards at the same time. By randomizing the range of each shard (within 0.75xSBS to 1.25xSBS), we ensure that nodes need to download shards at different times during training. If there is a cluster-level internet bandwidth limit, this PR will help alleviate potential bottlenecks.In this first image, we see that using the
py1br
shuffle algorithm can cause throughput drops when node internet bandwidth is very low. However, downloads are relatively balanced across nodes due to randomized shuffle block size.In this next image, we see that using the
py1e
shuffle algorithm eliminates throughput drops since downloads inside each node are much more balanced. However, overall downloads show a stepwise behavior since all nodes are downloading shards at the same time. This may cause issues with cluster-level internet bandwidth limits and cause unforeseen throughput drops due to this bottleneck.As a result, bringing the randomized shuffle block size from
py1br
topy1e
will ensure that the balanced intra-node downloading of the originalpy1e
is combined with the balanced inter-node downloading ofpy1br
.Issue #, if available:
Merge Checklist:
Put an
x
without space in the boxes that apply. If you are unsure about any checklist, please don't hesitate to ask. We are here to help! This is simply a reminder of what we are going to look for before merging your pull request.General
Tests
pre-commit
on my change. (check out thepre-commit
section of prerequisites)