[pytorch] process_group_agent optimizations #29324
Closed
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
This change adds a few optimizations to process_group_agent:
1) Don't add an extra tensor during serialization for the message id,
but instead put that in the preamble tensor. This saves maybe 15%
overhead for minimal-sized RPCs.
2) Add a payload-only fastpath.
a) For very tiny messages (e.g. acks, status updates), this reduces RPC overhead by roughly 50%, by removing zip file overhead and related setup/copying in torch::load()/torch::save().
b) If we do end up with large non-tensor payloads, this saves us ~25% in benchmarks, from avoiding copying in torch::save/load.
Differential Revision: [D18352261](https://our.internmc.facebook.com/intern/diff/D18352261/)
[ghstack-poisoned]
jjlilley
pushed a commit
that referenced
this pull request
Nov 6, 2019
This change adds a few optimizations to process_group_agent:
1) Don't add an extra tensor during serialization for the message id,
but instead put that in the preamble tensor. This saves maybe 15%
overhead for minimal-sized RPCs.
2) Add a payload-only fastpath.
a) For very tiny messages (e.g. acks, status updates), this reduces RPC overhead by roughly 50%, by removing zip file overhead and related setup/copying in torch::load()/torch::save().
b) If we do end up with large non-tensor payloads, this saves us ~25% in benchmarks, from avoiding copying in torch::save/load.
Differential Revision: [D18352261](https://our.internmc.facebook.com/intern/diff/D18352261/)
ghstack-source-id: 93389049
Pull Request resolved: #29324
mrshenli
reviewed
Nov 6, 2019
Contributor
mrshenli
left a comment
mrshenli
left a comment
There was a problem hiding this comment.
Question: if we switch from torch::save/torch::load to jit pickle/unpickle, do we still need to explicitly mark the serialization mode?
This change adds a few optimizations to process_group_agent:
1) Don't add an extra tensor during serialization for the message id,
but instead put that in the preamble tensor. This saves maybe 15%
overhead for minimal-sized RPCs.
2) Add a payload-only fastpath.
a) For very tiny messages (e.g. acks, status updates), this reduces RPC overhead by roughly 50%, by removing zip file overhead and related setup/copying in torch::load()/torch::save().
b) If we do end up with large non-tensor payloads, this saves us ~25% in benchmarks, from avoiding copying in torch::save/load.
Differential Revision: [D18352261](https://our.internmc.facebook.com/intern/diff/D18352261/)
[ghstack-poisoned]
jjlilley
pushed a commit
that referenced
this pull request
Nov 7, 2019
Pull Request resolved: #29324 This change adds a few optimizations to process_group_agent: 1) Don't add an extra tensor during serialization for the message id, but instead put that in the preamble tensor. This saves maybe 15% overhead for minimal-sized RPCs. 2) Add a payload-only fastpath. a) For very tiny messages (e.g. acks, status updates), this reduces RPC overhead by roughly 50%, by removing zip file overhead and related setup/copying in torch::load()/torch::save(). b) If we do end up with large non-tensor payloads, this saves us ~25% in benchmarks, from avoiding copying in torch::save/load. ghstack-source-id: 93413704 Differential Revision: [D18352261](https://our.internmc.facebook.com/intern/diff/D18352261/)
Author
If we decide that there's a better pickling mechanism, we can always get rid of this. My main goal here was that in case different sorts of encoding payloads have more optimal encoding schemes, that we have an out-of-band (and effectively free) mechanism to specify this. Right now, if we have a minimally-sized response from an RPC (e.g. an ACK), without any tensors, we expend more effort/bytes encoding this than we need to. Being able to alternate between torch::save / jit::pickle / payload-only / whatever, depending on the input, seems like it could help some cases. |
This change adds a few optimizations to process_group_agent:
1) Don't add an extra tensor during serialization for the message id,
but instead put that in the preamble tensor. This saves maybe 15%
overhead for minimal-sized RPCs.
2) Add a payload-only fastpath.
a) For very tiny messages (e.g. acks, status updates), this reduces RPC overhead by roughly 50%, by removing zip file overhead and related setup/copying in torch::load()/torch::save().
b) If we do end up with large non-tensor payloads, this saves us ~25% in benchmarks, from avoiding copying in torch::save/load.
Differential Revision: [D18352261](https://our.internmc.facebook.com/intern/diff/D18352261/)
[ghstack-poisoned]
jjlilley
pushed a commit
that referenced
this pull request
Nov 7, 2019
Pull Request resolved: #29324 This change adds a few optimizations to process_group_agent: 1) Don't add an extra tensor during serialization for the message id, but instead put that in the preamble tensor. This saves maybe 15% overhead for minimal-sized RPCs. 2) Add a payload-only fastpath. a) For very tiny messages (e.g. acks, status updates), this reduces RPC overhead by roughly 50%, by removing zip file overhead and related setup/copying in torch::load()/torch::save(). b) If we do end up with large non-tensor payloads, this saves us ~25% in benchmarks, from avoiding copying in torch::save/load. ghstack-source-id: 93489064 Differential Revision: [D18352261](https://our.internmc.facebook.com/intern/diff/D18352261/)
This change adds a few optimizations to process_group_agent:
1) Don't add an extra tensor during serialization for the message id,
but instead put that in the preamble tensor. This saves maybe 15%
overhead for minimal-sized RPCs.
2) Add a payload-only fastpath.
a) For very tiny messages (e.g. acks, status updates), this reduces RPC overhead by roughly 50%, by removing zip file overhead and related setup/copying in torch::load()/torch::save().
b) If we do end up with large non-tensor payloads, this saves us ~25% in benchmarks, from avoiding copying in torch::save/load.
Differential Revision: [D18352261](https://our.internmc.facebook.com/intern/diff/D18352261/)
[ghstack-poisoned]
This change adds a few optimizations to process_group_agent:
1) Don't add an extra tensor during serialization for the message id,
but instead put that in the preamble tensor. This saves maybe 15%
overhead for minimal-sized RPCs.
2) Add a payload-only fastpath.
a) For very tiny messages (e.g. acks, status updates), this reduces RPC overhead by roughly 50%, by removing zip file overhead and related setup/copying in torch::load()/torch::save().
b) If we do end up with large non-tensor payloads, this saves us ~25% in benchmarks, from avoiding copying in torch::save/load.
Differential Revision: [D18352261](https://our.internmc.facebook.com/intern/diff/D18352261/)
[ghstack-poisoned]
jjlilley
pushed a commit
that referenced
this pull request
Nov 8, 2019
Pull Request resolved: #29324 This change adds a few optimizations to process_group_agent: 1) Don't add an extra tensor during serialization for the message id, but instead put that in the preamble tensor. This saves maybe 15% overhead for minimal-sized RPCs. 2) Add a payload-only fastpath. a) For very tiny messages (e.g. acks, status updates), this reduces RPC overhead by roughly 50%, by removing zip file overhead and related setup/copying in torch::load()/torch::save(). b) If we do end up with large non-tensor payloads, this saves us ~25% in benchmarks, from avoiding copying in torch::save/load. ghstack-source-id: 93577550 Differential Revision: [D18352261](https://our.internmc.facebook.com/intern/diff/D18352261/)
Author
|
Closing this PR for a bit, there's another approach that I'd like to take... |
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:
This change adds a few optimizations to process_group_agent:
Don't add an extra tensor during serialization for the message id,
but instead put that in the preamble tensor. This saves maybe 15%
overhead for minimal-sized RPCs.
Add a payload-only fastpath.
a) For very tiny messages (e.g. acks, status updates), this reduces RPC overhead by roughly 50%, by removing zip file overhead and related setup/copying in torch::load()/torch::save().
b) If we do end up with large non-tensor payloads, this saves us ~25% in benchmarks, from avoiding copying in torch::save/load.
Differential Revision: D18352261