This repository has been archived by the owner on Jul 1, 2023. It is now read-only.
Add a Herring-like benchmark #410
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Summary: This diff adds two benchmarks that perform an allreduce operation using internally a technique inspired from Amazon's Herring paper. The two versions, one for TCP and one for InfiniBand interconnect, operate similarly: they spawn a set of clients (one per device, with multiple devices per machine, and multiple machines), then each allreduce call first performs a NCCL reduce or reduce_scatter step within each machine, then a shared send to a set of servers (one per machine, or one per device), then these servers perform the aggregation, and send back the result to the clients, which then do a final NCCL broadcast or all_gather step. Multiple allreduce calls, one per "bucket", are launched in parallel during each "epoch". There's also a baseline script that does these allreduces simply using NCCL. Since the benchmarks depend on PyTorch, they are a bit "awkward" to build and install from within the TensorPipe repo. So, instead of integrating them inside our CMake system, I opted to build them as a separate PyTorch extension, hence giving them a setup.py file. Thus the installation steps are to first install TensorPipe on its own, then NCCL, then PyTorch, and then the benchmarks. Differential Revision: D30220559 fbshipit-source-id: 43ab5bc440271942698313e6b06b39104a0e44de
|
This pull request was exported from Phabricator. Differential Revision: D30220559 |
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
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Outdated suggestions cannot be applied.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.
Summary:
This diff adds two benchmarks that perform an allreduce operation using internally a technique inspired from Amazon's Herring paper. The two versions, one for TCP and one for InfiniBand interconnect, operate similarly: they spawn a set of clients (one per device, with multiple devices per machine, and multiple machines), then each allreduce call first performs a NCCL reduce or reduce_scatter step within each machine, then a shared send to a set of servers (one per machine, or one per device), then these servers perform the aggregation, and send back the result to the clients, which then do a final NCCL broadcast or all_gather step. Multiple allreduce calls, one per "bucket", are launched in parallel during each "epoch". There's also a baseline script that does these allreduces simply using NCCL.
Since the benchmarks depend on PyTorch, they are a bit "awkward" to build and install from within the TensorPipe repo. So, instead of integrating them inside our CMake system, I opted to build them as a separate PyTorch extension, hence giving them a setup.py file. Thus the installation steps are to first install TensorPipe on its own, then NCCL, then PyTorch, and then the benchmarks.
Differential Revision: D30220559