[Profiler] Handle ABA for TensorImpl* when assigning IDs #87133
Closed
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
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/87133
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 FailuresAs of commit 06c42b2: The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Oct 17, 2022
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) [ghstack-poisoned]
robieta
pushed a commit
that referenced
this pull request
Oct 17, 2022
Pull Request resolved: #87133 Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. ghstack-source-id: 170664086 Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/)
slgong-fb
approved these changes
Oct 18, 2022
|
|
||
| for _ in range(3): | ||
| x.set_(torch.ones((1,)).storage()) | ||
| x.view_as(x) |
There was a problem hiding this comment.
Profiler only sees a Tensor (and therefore the (TensorImpl, StorageImpl) pair) when a Tensor is used in an op. Allocation just sees void*. So you have to do something with the Tensor to actually plumb it through an op, otherwise profiler will never see the new storage of x after the last set_. x.view_as(x) is basically a no-op. (I'll put that behind a staticmethod and write a good docstring.)
slgong-fb
reviewed
Oct 18, 2022
torch/csrc/profiler/collection.cpp
Outdated
| } | ||
|
|
||
| // Step 2) Handle the case that the storage of a TensorImpl changed. | ||
| // Step 5) Handle ABA for TensorImpl. |
There was a problem hiding this comment.
I really like this idea. I had a similar concern while writing a simplified logic for tensor identification/categorization.
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) [ghstack-poisoned]
robieta
pushed a commit
that referenced
this pull request
Oct 18, 2022
Pull Request resolved: #87133 Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. ghstack-source-id: 170713571 Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/)
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) [ghstack-poisoned]
robieta
pushed a commit
that referenced
this pull request
Oct 19, 2022
Pull Request resolved: #87133 Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. ghstack-source-id: 170863421 Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/)
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) [ghstack-poisoned]
robieta
added
release notes: profiler
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) [ghstack-poisoned]
robieta
pushed a commit
that referenced
this pull request
Oct 20, 2022
Pull Request resolved: #87133 Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. ghstack-source-id: 171000951 Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/)
added 2 commits
October 21, 2022 11:42
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) [ghstack-poisoned]
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) [ghstack-poisoned]
This was referenced Oct 23, 2022
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) [ghstack-poisoned]
robieta
pushed a commit
that referenced
this pull request
Oct 25, 2022
ID assignment has become a very complex facet of the profiler. The existing code has grown organically as I've discovered various refinements and has become very difficult to understand or reason about. (With more complexity coming in #87133) I want to take a step back and add some structure and additional comments to the ID assignment algorithm. Before I do, however, it's time to move it out of `collection.cpp` to a dedicated data flow file. Differential Revision: [D40666360](https://our.internmc.facebook.com/intern/diff/D40666360/) **NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D40666360/)! [ghstack-poisoned]
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) [ghstack-poisoned]
This was referenced Oct 25, 2022
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) [ghstack-poisoned]
robieta
pushed a commit
that referenced
this pull request
Oct 26, 2022
…e to `data_flow.cpp`" ID assignment has become a very complex facet of the profiler. The existing code has grown organically as I've discovered various refinements and has become very difficult to understand or reason about. (With more complexity coming in #87133) I want to take a step back and add some structure and additional comments to the ID assignment algorithm. Before I do, however, it's time to move it out of `collection.cpp` to a dedicated data flow file. Differential Revision: [D40666360](https://our.internmc.facebook.com/intern/diff/D40666360/) **NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D40666360/)! [ghstack-poisoned]
robieta
pushed a commit
that referenced
this pull request
Oct 26, 2022
…cpp`" ID assignment has become a very complex facet of the profiler. The existing code has grown organically as I've discovered various refinements and has become very difficult to understand or reason about. (With more complexity coming in #87133) I want to take a step back and add some structure and additional comments to the ID assignment algorithm. Before I do, however, it's time to move it out of `collection.cpp` to a dedicated data flow file. Differential Revision: [D40666360](https://our.internmc.facebook.com/intern/diff/D40666360/) **NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D40666360/)! [ghstack-poisoned]
pytorchmergebot
pushed a commit
that referenced
this pull request
Oct 28, 2022
ID assignment has become a very complex facet of the profiler. The existing code has grown organically as I've discovered various refinements and has become very difficult to understand or reason about. (With more complexity coming in #87133) I want to take a step back and add some structure and additional comments to the ID assignment algorithm. Before I do, however, it's time to move it out of `collection.cpp` to a dedicated data flow file. Differential Revision: [D40666360](https://our.internmc.facebook.com/intern/diff/D40666360/) **NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D40666360/)! Pull Request resolved: #87670 Approved by: https://github.com/slgong-fb
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) [ghstack-poisoned]
kulinseth
pushed a commit
to kulinseth/pytorch
that referenced
this pull request
Nov 5, 2022
…ch#87670) ID assignment has become a very complex facet of the profiler. The existing code has grown organically as I've discovered various refinements and has become very difficult to understand or reason about. (With more complexity coming in pytorch#87133) I want to take a step back and add some structure and additional comments to the ID assignment algorithm. Before I do, however, it's time to move it out of `collection.cpp` to a dedicated data flow file. Differential Revision: [D40666360](https://our.internmc.facebook.com/intern/diff/D40666360/) **NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D40666360/)! Pull Request resolved: pytorch#87670 Approved by: https://github.com/slgong-fb
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) [ghstack-poisoned]
chaekit
approved these changes
Nov 8, 2022
kulinseth
pushed a commit
to kulinseth/pytorch
that referenced
this pull request
Dec 10, 2022
…ch#87670) ID assignment has become a very complex facet of the profiler. The existing code has grown organically as I've discovered various refinements and has become very difficult to understand or reason about. (With more complexity coming in pytorch#87133) I want to take a step back and add some structure and additional comments to the ID assignment algorithm. Before I do, however, it's time to move it out of `collection.cpp` to a dedicated data flow file. Differential Revision: [D40666360](https://our.internmc.facebook.com/intern/diff/D40666360/) **NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D40666360/)! Pull Request resolved: pytorch#87670 Approved by: https://github.com/slgong-fb
kulinseth
pushed a commit
to kulinseth/pytorch
that referenced
this pull request
Dec 10, 2022
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem) ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead. This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.) The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues. Differential Revision: [D40445121](https://our.internmc.facebook.com/intern/diff/D40445121/) Pull Request resolved: pytorch#87133 Approved by: https://github.com/slgong-fb, https://github.com/chaekit
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 (oldest at bottom):
Part of the current ID assingment algorithm groups any Storages which are associated with the same TensorImpl*. This isn't sound (which I knew but deferred until it actually became a problem) because pointers can be reused by different objects. (ABA problem)
ABA is easy to handle for Storage because we see allocations and frees, but ~TensorImpl is very hot and cannot tolerate profiling code without significant increases in overhead.
This PR narrows the conditions under which ID assignment will join on TensorImpl*. Two storages which are associated with the same TensorImpl* are grouped IFF they were live at the same time. (Note that this still allows storages with disjoint lifetimes to be joined transitively through a third storage which overlaps with both.)
The need for this PR arose in memory profiling. The Python argument parser creates short lived Tensors for (some) scalar arguments which triggers this issue. (Which is stochastic and platform dependent since optimizations like reusing recently freed allocations is implementation defined.) Spurious connections can lead to confusing and long range interactions when building up the memory profile, so it makes sense to harden ID assignment to avoid any issues.
Differential Revision: D40445121