[NCCL] DDP communication hook: getFuture() without cudaStreamAddCallback #42335
Conversation
This file contains hidden or 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
**Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). See the following approach: **New design:** 1. As we create a `WorkNCCL` object inside `ProcessGroupNCCL::collective`, we also create a `FutureNCCL` (sub class of `Future`) associated with `WorkNCCL` and store it with the object. This future has a reference to the original work object. 2. The Future is marked as completed when its created (to allow for async execution of callbacks). 3. fut.wait() simply synchronizes the streams (synchronizeStreams() in Sinan's diff) and returns the result. This preserves the async execution model for CUDA. 4. When we add a callback to this future (.then()), we synchronizeStreams() and invoke the callback inline. Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/) [ghstack-poisoned]
sinannasir
requested review from
apaszke,
mrshenli,
pietern,
pritamdamania87 and
zhaojuanmao
as code owners
sinannasir
added a commit
that referenced
this pull request
Jul 30, 2020
**Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). See the following approach: **New design:** 1. As we create a `WorkNCCL` object inside `ProcessGroupNCCL::collective`, we also create a `FutureNCCL` (sub class of `Future`) associated with `WorkNCCL` and store it with the object. This future has a reference to the original work object. 2. The Future is marked as completed when its created (to allow for async execution of callbacks). 3. fut.wait() simply synchronizes the streams (synchronizeStreams() in Sinan's diff) and returns the result. This preserves the async execution model for CUDA. 4. When we add a callback to this future (.then()), we synchronizeStreams() and invoke the callback inline. Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/) ghstack-source-id: 108894839 Pull Request resolved: #42335
💊 CI failures summary and remediationsAs of commit 6322122 (more details on the Dr. CI page):
ci.pytorch.org: 1 failedThis comment was automatically generated by Dr. CI (expand for details).Follow this link to opt-out of these comments for your Pull Requests.Please report bugs/suggestions on the GitHub issue tracker or post in the (internal) Dr. CI Users group. This comment has been revised 63 times. |
…eamAddCallback" **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). See the following approach: **New design:** 1. As we create a `WorkNCCL` object inside `ProcessGroupNCCL::collective`, we also create a `FutureNCCL` (sub class of `Future`) associated with `WorkNCCL` and store it with the object. This future has a reference to the original work object. 2. The Future is marked as completed when its created (to allow for async execution of callbacks). 3. fut.wait() simply synchronizes the streams (synchronizeStreams() in Sinan's diff) and returns the result. This preserves the async execution model for CUDA. 4. When we add a callback to this future (.then()), we synchronizeStreams() and invoke the callback inline. Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/) [ghstack-poisoned]
…eamAddCallback" **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/) [ghstack-poisoned]
sinannasir
added a commit
that referenced
this pull request
Jul 31, 2020
Pull Request resolved: #42335 **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). ghstack-source-id: 108915990 Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/)
pritamdamania87
suggested changes
Jul 31, 2020
…eamAddCallback" **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/) [ghstack-poisoned]
sinannasir
added a commit
that referenced
this pull request
Jul 31, 2020
Pull Request resolved: #42335 **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). ghstack-source-id: 108944549 Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/)
mrshenli
reviewed
Jul 31, 2020
mrshenli
reviewed
Jul 31, 2020
mrshenli
reviewed
Jul 31, 2020
mrshenli
reviewed
Jul 31, 2020
mrshenli
reviewed
Jul 31, 2020
mrshenli
requested changes
Jul 31, 2020
mrshenli
reviewed
Jul 31, 2020
…eamAddCallback" **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/) [ghstack-poisoned]
sinannasir
added a commit
that referenced
this pull request
Jul 31, 2020
Pull Request resolved: #42335 **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). ghstack-source-id: 108986334 Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/)
rohan-varma
reviewed
Aug 1, 2020
rohan-varma
reviewed
Aug 1, 2020
…eamAddCallback" **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/) [ghstack-poisoned]
sinannasir
added a commit
that referenced
this pull request
Aug 4, 2020
Pull Request resolved: #42335 **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). ghstack-source-id: 109150639 Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/)
…eamAddCallback" **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/) [ghstack-poisoned]
sinannasir
added a commit
that referenced
this pull request
Aug 4, 2020
Pull Request resolved: #42335 **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). ghstack-source-id: 109156649 Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/)
mrshenli
reviewed
Aug 4, 2020
…eamAddCallback" **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/) [ghstack-poisoned]
sinannasir
added a commit
that referenced
this pull request
Aug 5, 2020
Pull Request resolved: #42335 **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). ghstack-source-id: 109225244 Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/)
…eamAddCallback" **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/) [ghstack-poisoned]
sinannasir
added a commit
that referenced
this pull request
Aug 5, 2020
Pull Request resolved: #42335 **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). ghstack-source-id: 109291702 Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/)
…eamAddCallback" **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/) [ghstack-poisoned]
sinannasir
added a commit
that referenced
this pull request
Aug 6, 2020
Pull Request resolved: #42335 **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). ghstack-source-id: 109392464 Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/)
mrshenli
approved these changes
Aug 7, 2020
There was a problem hiding this comment.
Thanks for fixing the circular reference. LGTM! Please address @pritamdamania87's comments before landing. Thanks!
…eamAddCallback" **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/) [ghstack-poisoned]
sinannasir
added a commit
that referenced
this pull request
Aug 7, 2020
Pull Request resolved: #42335 **Main goal:** For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff. We add an API `c10::intrusive_ptr<c10::ivalue::Future> getFuture()` to `c10d::ProcessGroup::Work`. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation. We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in [#41596](#41596). ghstack-source-id: 109461507 Differential Revision: [D22833298](https://our.internmc.facebook.com/intern/diff/D22833298/)
|
This pull request has been merged in 0a804be. |
3 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Stack from ghstack:
workandfuture_work#41840 [NCCL] [For Test] In DDP's reducer mergeworkandfuture_workMain goal: For DDP communication hook, provide an API called "get_future" to retrieve a future associated with the completion of c10d.ProcessGroupNCCL.work. Enable NCCL support for this API in this diff.
We add an API
c10::intrusive_ptr<c10::ivalue::Future> getFuture()toc10d::ProcessGroup::Work. This API will only be supported by NCCL in the first version, the default implementation will throw UnsupportedOperation.We no longer consider a design that involves cudaStreamAddCallback which potentially was causing performance regression in #41596.
Differential Revision: D22833298