Add cache conflict miss support (backend) #2596
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
|
This pull request was exported from Phabricator. Differential Revision: D55998215 |
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
|
This pull request was exported from Phabricator. Differential Revision: D55998215 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
May 20, 2024
Summary: Pull Request resolved: pytorch#2596 Differential Revision: D55998215
|
This pull request was exported from Phabricator. Differential Revision: D55998215 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
May 20, 2024
Summary: Pull Request resolved: pytorch#2596 Differential Revision: D55998215
|
This pull request was exported from Phabricator. Differential Revision: D55998215 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
May 20, 2024
Summary: Pull Request resolved: pytorch#2596 Differential Revision: D55998215
|
This pull request was exported from Phabricator. Differential Revision: D55998215 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
May 22, 2024
Summary: Pull Request resolved: pytorch#2596 Differential Revision: D55998215
|
This pull request was exported from Phabricator. Differential Revision: D55998215 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
May 22, 2024
Summary: Pull Request resolved: pytorch#2596 Prior to this diff, SSD TBE lacked support for the conflict cache miss scenario. It operated under the assumption that the cache, located in GPU memory, was sufficiently large to hold all prefetched data from SSD. In the event of a conflict cache miss, the behavior of SSD TBE would be unpredictable (it could either fail or potentially access illegal memory). Note that a conflict cache miss happens when an embedding row is absent in the cache, and after being fetched from SSD, it cannot be inserted into the cache due to capacity constraints or associativity limitations. This diff introduces support for conflict cache misses by storing rows that cannot be inserted into the cache due to conflicts in a scratch pad, which is a temporary GPU tensor. In the case where rows are missed from the cache, TBE kernels can access the scratch pad. Prior to this diff, during the SSD prefetch stage, any row that was missed the cache and required fetching from SSD would be first fetched into a CPU scratch pad and then transferred to GPU. Rows that could be inserted into the cache would subsequently be copied from the GPU scratch pad into the cache. If conflict misses occurred, the prefetch behavior would be unpredictable. With this diff, conflict missed rows are now retained in the scratch pad, which is kept alive until the current iteration completes. Throughout the forward and backward + optimizer stages of TBE, both the cache and scratch pad are equivalent in terms of usage. However, following the completion of the backward + optimizer step, rows in the scratch pad are flushed back to SSD, unlike rows residing in the cache which are not evicted for future usage (see the diagram below for more details). {F1645878181} Differential Revision: D55998215
|
This pull request was exported from Phabricator. Differential Revision: D55998215 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
May 22, 2024
Summary: Pull Request resolved: pytorch#2596 Prior to this diff, SSD TBE lacked support for the conflict cache miss scenario. It operated under the assumption that the cache, located in GPU memory, was sufficiently large to hold all prefetched data from SSD. In the event of a conflict cache miss, the behavior of SSD TBE would be unpredictable (it could either fail or potentially access illegal memory). Note that a conflict cache miss happens when an embedding row is absent in the cache, and after being fetched from SSD, it cannot be inserted into the cache due to capacity constraints or associativity limitations. This diff introduces support for conflict cache misses by storing rows that cannot be inserted into the cache due to conflicts in a scratch pad, which is a temporary GPU tensor. In the case where rows are missed from the cache, TBE kernels can access the scratch pad. Prior to this diff, during the SSD prefetch stage, any row that was missed the cache and required fetching from SSD would be first fetched into a CPU scratch pad and then transferred to GPU. Rows that could be inserted into the cache would subsequently be copied from the GPU scratch pad into the cache. If conflict misses occurred, the prefetch behavior would be unpredictable. With this diff, conflict missed rows are now retained in the scratch pad, which is kept alive until the current iteration completes. Throughout the forward and backward + optimizer stages of TBE, both the cache and scratch pad are equivalent in terms of usage. However, following the completion of the backward + optimizer step, rows in the scratch pad are flushed back to SSD, unlike rows residing in the cache which are not evicted for future usage (see the diagram below for more details). {F1645878181} Differential Revision: D55998215
Summary: Pull Request resolved: pytorch#2596 Prior to this diff, SSD TBE lacked support for the conflict cache miss scenario. It operated under the assumption that the cache, located in GPU memory, was sufficiently large to hold all prefetched data from SSD. In the event of a conflict cache miss, the behavior of SSD TBE would be unpredictable (it could either fail or potentially access illegal memory). Note that a conflict cache miss happens when an embedding row is absent in the cache, and after being fetched from SSD, it cannot be inserted into the cache due to capacity constraints or associativity limitations. This diff introduces support for conflict cache misses by storing rows that cannot be inserted into the cache due to conflicts in a scratch pad, which is a temporary GPU tensor. In the case where rows are missed from the cache, TBE kernels can access the scratch pad. Prior to this diff, during the SSD prefetch stage, any row that was missed the cache and required fetching from SSD would be first fetched into a CPU scratch pad and then transferred to GPU. Rows that could be inserted into the cache would subsequently be copied from the GPU scratch pad into the cache. If conflict misses occurred, the prefetch behavior would be unpredictable. With this diff, conflict missed rows are now retained in the scratch pad, which is kept alive until the current iteration completes. Throughout the forward and backward + optimizer stages of TBE, both the cache and scratch pad are equivalent in terms of usage. However, following the completion of the backward + optimizer step, rows in the scratch pad are flushed back to SSD, unlike rows residing in the cache which are not evicted for future usage (see the diagram below for more details). {F1645878181} Differential Revision: D55998215
|
This pull request was exported from Phabricator. Differential Revision: D55998215 |
joshuadeng
pushed a commit
to joshuadeng/FBGEMM
that referenced
this pull request
May 23, 2024
Summary: Pull Request resolved: pytorch#2596 Prior to this diff, SSD TBE lacked support for the conflict cache miss scenario. It operated under the assumption that the cache, located in GPU memory, was sufficiently large to hold all prefetched data from SSD. In the event of a conflict cache miss, the behavior of SSD TBE would be unpredictable (it could either fail or potentially access illegal memory). Note that a conflict cache miss happens when an embedding row is absent in the cache, and after being fetched from SSD, it cannot be inserted into the cache due to capacity constraints or associativity limitations. This diff introduces support for conflict cache misses by storing rows that cannot be inserted into the cache due to conflicts in a scratch pad, which is a temporary GPU tensor. In the case where rows are missed from the cache, TBE kernels can access the scratch pad. Prior to this diff, during the SSD prefetch stage, any row that was missed the cache and required fetching from SSD would be first fetched into a CPU scratch pad and then transferred to GPU. Rows that could be inserted into the cache would subsequently be copied from the GPU scratch pad into the cache. If conflict misses occurred, the prefetch behavior would be unpredictable. With this diff, conflict missed rows are now retained in the scratch pad, which is kept alive until the current iteration completes. Throughout the forward and backward + optimizer stages of TBE, both the cache and scratch pad are equivalent in terms of usage. However, following the completion of the backward + optimizer step, rows in the scratch pad are flushed back to SSD, unlike rows residing in the cache which are not evicted for future usage (see the diagram below for more details). {F1645878181} Differential Revision: D55998215
joshuadeng
pushed a commit
to joshuadeng/FBGEMM
that referenced
this pull request
May 23, 2024
Summary: Pull Request resolved: pytorch#2596 Prior to this diff, SSD TBE lacked support for the conflict cache miss scenario. It operated under the assumption that the cache, located in GPU memory, was sufficiently large to hold all prefetched data from SSD. In the event of a conflict cache miss, the behavior of SSD TBE would be unpredictable (it could either fail or potentially access illegal memory). Note that a conflict cache miss happens when an embedding row is absent in the cache, and after being fetched from SSD, it cannot be inserted into the cache due to capacity constraints or associativity limitations. This diff introduces support for conflict cache misses by storing rows that cannot be inserted into the cache due to conflicts in a scratch pad, which is a temporary GPU tensor. In the case where rows are missed from the cache, TBE kernels can access the scratch pad. Prior to this diff, during the SSD prefetch stage, any row that was missed the cache and required fetching from SSD would be first fetched into a CPU scratch pad and then transferred to GPU. Rows that could be inserted into the cache would subsequently be copied from the GPU scratch pad into the cache. If conflict misses occurred, the prefetch behavior would be unpredictable. With this diff, conflict missed rows are now retained in the scratch pad, which is kept alive until the current iteration completes. Throughout the forward and backward + optimizer stages of TBE, both the cache and scratch pad are equivalent in terms of usage. However, following the completion of the backward + optimizer step, rows in the scratch pad are flushed back to SSD, unlike rows residing in the cache which are not evicted for future usage (see the diagram below for more details). {F1645878181} Differential Revision: D55998215
joshuadeng
pushed a commit
to joshuadeng/FBGEMM
that referenced
this pull request
May 24, 2024
Summary: Pull Request resolved: pytorch#2596 Prior to this diff, SSD TBE lacked support for the conflict cache miss scenario. It operated under the assumption that the cache, located in GPU memory, was sufficiently large to hold all prefetched data from SSD. In the event of a conflict cache miss, the behavior of SSD TBE would be unpredictable (it could either fail or potentially access illegal memory). Note that a conflict cache miss happens when an embedding row is absent in the cache, and after being fetched from SSD, it cannot be inserted into the cache due to capacity constraints or associativity limitations. This diff introduces support for conflict cache misses by storing rows that cannot be inserted into the cache due to conflicts in a scratch pad, which is a temporary GPU tensor. In the case where rows are missed from the cache, TBE kernels can access the scratch pad. Prior to this diff, during the SSD prefetch stage, any row that was missed the cache and required fetching from SSD would be first fetched into a CPU scratch pad and then transferred to GPU. Rows that could be inserted into the cache would subsequently be copied from the GPU scratch pad into the cache. If conflict misses occurred, the prefetch behavior would be unpredictable. With this diff, conflict missed rows are now retained in the scratch pad, which is kept alive until the current iteration completes. Throughout the forward and backward + optimizer stages of TBE, both the cache and scratch pad are equivalent in terms of usage. However, following the completion of the backward + optimizer step, rows in the scratch pad are flushed back to SSD, unlike rows residing in the cache which are not evicted for future usage (see the diagram below for more details). {F1645878181} Differential Revision: D55998215
|
This pull request has been merged in db4d379. |
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.
Differential Revision: D55998215