Implement multi-pass prefetch for memory efficiency by levythu · Pull Request #2566 · pytorch/FBGEMM

levythu · 2024-05-07T15:27:33Z

Summary:

Context

Memory snapshot shows significant memory usage during prefetch kernels (specifically, linearize_cache_index and lru_cache_populate), which is estimated to be 6x of input size

And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty.

So we need to lower down the peak prefetch memory usage as much as possible.

MultiPass Prefetch (MPP)

Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is O(N), so we reduce the total prefetched index (N) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint.

Benefit

With this being turned on, the peak memory usage will be dropped from 6 * input_size to (6 / M) * input_size, where M is the total # of passes being configured.

Overhead
Overall, the bigger M we configured, the slower we'll be. But the overall overhead is acceptable.

Efficiency regression: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes.
- The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes.
- The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible.
Spamming CUDA Launch Queue: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass.

What's in the patch?

Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility
Modify the lru_find_uncached to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one.

Differential Revision: D56908989

facebook-github-bot · 2024-05-07T15:27:39Z

This pull request was exported from Phabricator. Differential Revision: D56908989

netlify · 2024-05-07T15:27:47Z

❌ Deploy Preview for pytorch-fbgemm-docs failed.

Name	Link
🔨 Latest commit	`3386341`
🔍 Latest deploy log	https://app.netlify.com/sites/pytorch-fbgemm-docs/deploys/66469ed5af92290008b814cd

Summary: ## Context Memory snapshot shows significant memory usage during prefetch kernels (specifically, `linearize_cache_index` and `lru_cache_populate`), which is estimated to be 6x of input size And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty. So we need to lower down the peak prefetch memory usage as much as possible. ## MultiPass Prefetch (MPP) Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is `O(N)`, so we reduce the total prefetched index (`N`) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint. **Benefit** With this being turned on, the peak memory usage will be dropped from `6 * input_size` to `(6 / M) * input_size`, where `M` is the total # of passes being configured. **Overhead** Overall, the bigger `M` we configured, the slower we'll be. But the overall overhead is acceptable. - **Efficiency regression**: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes. - The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes. - The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible. - **Spamming CUDA Launch Queue**: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass. ## What's in the patch? 1. Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility 2. Modify the `lru_find_uncached` to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one. Differential Revision: D56908989

facebook-github-bot · 2024-05-07T15:28:53Z

This pull request was exported from Phabricator. Differential Revision: D56908989

…r will correctly propagate its storage cost Summary: After FBGEMM TBE support multipass prefetch mode (see pytorch/FBGEMM#2566 for the full context), this diff will enable TorchRec to pass it all through via CacheParams, and shard estimator will recognize the memory saving accordingly. Differential Revision: D57055184

…r will correctly propagate its storage cost (meta-pytorch#2000) Summary: After FBGEMM TBE support multipass prefetch mode (see pytorch/FBGEMM#2566 for the full context), this diff will enable TorchRec to pass it all through via CacheParams, and shard estimator will recognize the memory saving accordingly. Differential Revision: D57055184

Summary: ## Context Memory snapshot shows significant memory usage during prefetch kernels (specifically, `linearize_cache_index` and `lru_cache_populate`), which is estimated to be 6x of input size And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty. So we need to lower down the peak prefetch memory usage as much as possible. ## MultiPass Prefetch (MPP) Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is `O(N)`, so we reduce the total prefetched index (`N`) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint. **Benefit** With this being turned on, the peak memory usage will be dropped from `6 * input_size` to `(6 / M) * input_size`, where `M` is the total # of passes being configured. **Overhead** Overall, the bigger `M` we configured, the slower we'll be. But the overall overhead is acceptable. - **Efficiency regression**: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes. - The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes. - The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible. - **Spamming CUDA Launch Queue**: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass. ## What's in the patch? 1. Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility 2. Modify the `lru_find_uncached` to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one. Reviewed By: sryap Differential Revision: D56908989

facebook-github-bot · 2024-05-15T18:22:02Z

This pull request was exported from Phabricator. Differential Revision: D56908989

Summary: ## Context Memory snapshot shows significant memory usage during prefetch kernels (specifically, `linearize_cache_index` and `lru_cache_populate`), which is estimated to be 6x of input size And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty. So we need to lower down the peak prefetch memory usage as much as possible. ## MultiPass Prefetch (MPP) Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is `O(N)`, so we reduce the total prefetched index (`N`) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint. **Benefit** With this being turned on, the peak memory usage will be dropped from `6 * input_size` to `(6 / M) * input_size`, where `M` is the total # of passes being configured. **Overhead** Overall, the bigger `M` we configured, the slower we'll be. But the overall overhead is acceptable. - **Efficiency regression**: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes. - The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes. - The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible. - **Spamming CUDA Launch Queue**: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass. ## What's in the patch? 1. Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility 2. Modify the `lru_find_uncached` to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one. Reviewed By: sryap Differential Revision: D56908989

facebook-github-bot · 2024-05-15T20:42:35Z

This pull request was exported from Phabricator. Differential Revision: D56908989

Summary: ## Context Memory snapshot shows significant memory usage during prefetch kernels (specifically, `linearize_cache_index` and `lru_cache_populate`), which is estimated to be 6x of input size And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty. So we need to lower down the peak prefetch memory usage as much as possible. ## MultiPass Prefetch (MPP) Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is `O(N)`, so we reduce the total prefetched index (`N`) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint. **Benefit** With this being turned on, the peak memory usage will be dropped from `6 * input_size` to `(6 / M) * input_size`, where `M` is the total # of passes being configured. **Overhead** Overall, the bigger `M` we configured, the slower we'll be. But the overall overhead is acceptable. - **Efficiency regression**: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes. - The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes. - The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible. - **Spamming CUDA Launch Queue**: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass. ## What's in the patch? 1. Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility 2. Modify the `lru_find_uncached` to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one. Reviewed By: sryap Differential Revision: D56908989

facebook-github-bot · 2024-05-15T20:43:15Z

This pull request was exported from Phabricator. Differential Revision: D56908989

Summary: ## Context Memory snapshot shows significant memory usage during prefetch kernels (specifically, `linearize_cache_index` and `lru_cache_populate`), which is estimated to be 6x of input size And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty. So we need to lower down the peak prefetch memory usage as much as possible. ## MultiPass Prefetch (MPP) Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is `O(N)`, so we reduce the total prefetched index (`N`) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint. **Benefit** With this being turned on, the peak memory usage will be dropped from `6 * input_size` to `(6 / M) * input_size`, where `M` is the total # of passes being configured. **Overhead** Overall, the bigger `M` we configured, the slower we'll be. But the overall overhead is acceptable. - **Efficiency regression**: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes. - The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes. - The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible. - **Spamming CUDA Launch Queue**: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass. ## What's in the patch? 1. Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility 2. Modify the `lru_find_uncached` to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one. Reviewed By: sryap Differential Revision: D56908989

facebook-github-bot · 2024-05-15T21:56:07Z

This pull request was exported from Phabricator. Differential Revision: D56908989

Summary: ## Context Memory snapshot shows significant memory usage during prefetch kernels (specifically, `linearize_cache_index` and `lru_cache_populate`), which is estimated to be 6x of input size And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty. So we need to lower down the peak prefetch memory usage as much as possible. ## MultiPass Prefetch (MPP) Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is `O(N)`, so we reduce the total prefetched index (`N`) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint. **Benefit** With this being turned on, the peak memory usage will be dropped from `6 * input_size` to `(6 / M) * input_size`, where `M` is the total # of passes being configured. **Overhead** Overall, the bigger `M` we configured, the slower we'll be. But the overall overhead is acceptable. - **Efficiency regression**: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes. - The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes. - The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible. - **Spamming CUDA Launch Queue**: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass. ## What's in the patch? 1. Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility 2. Modify the `lru_find_uncached` to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one. Reviewed By: sryap Differential Revision: D56908989

facebook-github-bot · 2024-05-16T02:53:43Z

This pull request was exported from Phabricator. Differential Revision: D56908989

Summary: ## Context Memory snapshot shows significant memory usage during prefetch kernels (specifically, `linearize_cache_index` and `lru_cache_populate`), which is estimated to be 6x of input size And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty. So we need to lower down the peak prefetch memory usage as much as possible. ## MultiPass Prefetch (MPP) Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is `O(N)`, so we reduce the total prefetched index (`N`) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint. **Benefit** With this being turned on, the peak memory usage will be dropped from `6 * input_size` to `(6 / M) * input_size`, where `M` is the total # of passes being configured. **Overhead** Overall, the bigger `M` we configured, the slower we'll be. But the overall overhead is acceptable. - **Efficiency regression**: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes. - The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes. - The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible. - **Spamming CUDA Launch Queue**: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass. ## What's in the patch? 1. Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility 2. Modify the `lru_find_uncached` to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one. Reviewed By: sryap Differential Revision: D56908989

facebook-github-bot · 2024-05-16T02:55:04Z

This pull request was exported from Phabricator. Differential Revision: D56908989

Summary: ## Context Memory snapshot shows significant memory usage during prefetch kernels (specifically, `linearize_cache_index` and `lru_cache_populate`), which is estimated to be 6x of input size And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty. So we need to lower down the peak prefetch memory usage as much as possible. ## MultiPass Prefetch (MPP) Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is `O(N)`, so we reduce the total prefetched index (`N`) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint. **Benefit** With this being turned on, the peak memory usage will be dropped from `6 * input_size` to `(6 / M) * input_size`, where `M` is the total # of passes being configured. **Overhead** Overall, the bigger `M` we configured, the slower we'll be. But the overall overhead is acceptable. - **Efficiency regression**: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes. - The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes. - The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible. - **Spamming CUDA Launch Queue**: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass. ## What's in the patch? 1. Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility 2. Modify the `lru_find_uncached` to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one. Reviewed By: sryap Differential Revision: D56908989

facebook-github-bot · 2024-05-16T15:02:46Z

This pull request was exported from Phabricator. Differential Revision: D56908989

Summary: ## Context Memory snapshot shows significant memory usage during prefetch kernels (specifically, `linearize_cache_index` and `lru_cache_populate`), which is estimated to be 6x of input size And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty. So we need to lower down the peak prefetch memory usage as much as possible. ## MultiPass Prefetch (MPP) Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is `O(N)`, so we reduce the total prefetched index (`N`) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint. **Benefit** With this being turned on, the peak memory usage will be dropped from `6 * input_size` to `(6 / M) * input_size`, where `M` is the total # of passes being configured. **Overhead** Overall, the bigger `M` we configured, the slower we'll be. But the overall overhead is acceptable. - **Efficiency regression**: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes. - The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes. - The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible. - **Spamming CUDA Launch Queue**: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass. ## What's in the patch? 1. Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility 2. Modify the `lru_find_uncached` to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one. Reviewed By: sryap Differential Revision: D56908989

facebook-github-bot · 2024-05-16T15:03:13Z

This pull request was exported from Phabricator. Differential Revision: D56908989

Summary: ## Context Memory snapshot shows significant memory usage during prefetch kernels (specifically, `linearize_cache_index` and `lru_cache_populate`), which is estimated to be 6x of input size And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty. So we need to lower down the peak prefetch memory usage as much as possible. ## MultiPass Prefetch (MPP) Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is `O(N)`, so we reduce the total prefetched index (`N`) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint. **Benefit** With this being turned on, the peak memory usage will be dropped from `6 * input_size` to `(6 / M) * input_size`, where `M` is the total # of passes being configured. **Overhead** Overall, the bigger `M` we configured, the slower we'll be. But the overall overhead is acceptable. - **Efficiency regression**: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes. - The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes. - The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible. - **Spamming CUDA Launch Queue**: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass. ## What's in the patch? 1. Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility 2. Modify the `lru_find_uncached` to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one. Reviewed By: sryap Differential Revision: D56908989

facebook-github-bot · 2024-05-16T15:14:08Z

This pull request was exported from Phabricator. Differential Revision: D56908989

Summary: ## Context Memory snapshot shows significant memory usage during prefetch kernels (specifically, `linearize_cache_index` and `lru_cache_populate`), which is estimated to be 6x of input size And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty. So we need to lower down the peak prefetch memory usage as much as possible. ## MultiPass Prefetch (MPP) Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is `O(N)`, so we reduce the total prefetched index (`N`) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint. **Benefit** With this being turned on, the peak memory usage will be dropped from `6 * input_size` to `(6 / M) * input_size`, where `M` is the total # of passes being configured. **Overhead** Overall, the bigger `M` we configured, the slower we'll be. But the overall overhead is acceptable. - **Efficiency regression**: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes. - The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes. - The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible. - **Spamming CUDA Launch Queue**: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass. ## What's in the patch? 1. Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility 2. Modify the `lru_find_uncached` to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one. Reviewed By: sryap Differential Revision: D56908989

facebook-github-bot · 2024-05-16T15:15:15Z

This pull request was exported from Phabricator. Differential Revision: D56908989

…r will correctly propagate its storage cost (meta-pytorch#2000) Summary: After FBGEMM TBE support multipass prefetch mode (see pytorch/FBGEMM#2566 for the full context), this diff will enable TorchRec to pass it all through via CacheParams, and shard estimator will recognize the memory saving accordingly. Differential Revision: D57055184

Summary: ## Context Memory snapshot shows significant memory usage during prefetch kernels (specifically, `linearize_cache_index` and `lru_cache_populate`), which is estimated to be 6x of input size And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty. So we need to lower down the peak prefetch memory usage as much as possible. ## MultiPass Prefetch (MPP) Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is `O(N)`, so we reduce the total prefetched index (`N`) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint. **Benefit** With this being turned on, the peak memory usage will be dropped from `6 * input_size` to `(6 / M) * input_size`, where `M` is the total # of passes being configured. **Overhead** Overall, the bigger `M` we configured, the slower we'll be. But the overall overhead is acceptable. - **Efficiency regression**: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes. - The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes. - The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible. - **Spamming CUDA Launch Queue**: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass. ## What's in the patch? 1. Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility 2. Modify the `lru_find_uncached` to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one. Reviewed By: sryap Differential Revision: D56908989

facebook-github-bot · 2024-05-16T19:34:57Z

This pull request was exported from Phabricator. Differential Revision: D56908989

Summary: ## Context Memory snapshot shows significant memory usage during prefetch kernels (specifically, `linearize_cache_index` and `lru_cache_populate`), which is estimated to be 6x of input size And unfortunately, due to they using dedicated stream, the memory cannot be reused by any other stream without performance penalty. So we need to lower down the peak prefetch memory usage as much as possible. ## MultiPass Prefetch (MPP) Multipass prefetch is basically a technique to sacrifice a bit of more running time for less peak memory during prefetch: We observed that intermediate memory usage for all functions during prefetch is `O(N)`, so we reduce the total prefetched index (`N`) for each pass to reduce the peak temporary usage. The following passes will recycle the memory used in the first pass so they won't further increase the memory footprint. **Benefit** With this being turned on, the peak memory usage will be dropped from `6 * input_size` to `(6 / M) * input_size`, where `M` is the total # of passes being configured. **Overhead** Overall, the bigger `M` we configured, the slower we'll be. But the overall overhead is acceptable. - **Efficiency regression**: Prefetch is taking longer because the process of cache lookup is being repeated for every duplicate index. In the past, they're deduped before being looked up, but now they might be look up multiple times if duplicate index are across different passes. - The regression is overall insignificant, as the major cost is the data movement between DDR and HBM. We'll always copy the data only once, even if they're duplicated across different passes. - The regression is likely hidden from the actual training performance, since prefetch happen in a separate stream. As long as it's not long enough to block sparse backward it's invisible. - **Spamming CUDA Launch Queue**: CUDA is allowing max # of 1024 pending kernels. CPU will go blocking if more are submitted. If a kernel is really small, we'll easily spam launch queue and greatly hurt QPS. We mitigate this via limit the minimal # of elements for a pass. ## What's in the patch? 1. Add multipass prefetch config to the interface of TBE. By default it's None for full backward compatibility 2. Modify the `lru_find_uncached` to make it idempotent -- if we tried to lock the same id multiple times in one single timestep (but multiple passes), we'll increase lock counter by only one. Reviewed By: sryap Differential Revision: D56908989

facebook-github-bot · 2024-05-17T00:03:39Z

This pull request was exported from Phabricator. Differential Revision: D56908989

facebook-github-bot · 2024-05-17T00:03:42Z

This pull request was exported from Phabricator. Differential Revision: D56908989

facebook-github-bot · 2024-05-17T14:37:12Z

This pull request has been merged in 578ab67.

…r will correctly propagate its storage cost (meta-pytorch#2000) Summary: After FBGEMM TBE support multipass prefetch mode (see pytorch/FBGEMM#2566 for the full context), this diff will enable TorchRec to pass it all through via CacheParams, and shard estimator will recognize the memory saving accordingly. Reviewed By: sarckk Differential Revision: D57055184

…r will correctly propagate its storage cost (#2000) Summary: Pull Request resolved: #2000 After FBGEMM TBE support multipass prefetch mode (see pytorch/FBGEMM#2566 for the full context), this diff will enable TorchRec to pass it all through via CacheParams, and shard estimator will recognize the memory saving accordingly. Reviewed By: sarckk Differential Revision: D57055184 fbshipit-source-id: 9fc2d05c30d7826f654976b1d85059fb3e9a1aae

facebook-github-bot added cla signed fb-exported labels May 7, 2024

levythu force-pushed the export-D56908989 branch from 133e6dc to 48773b2 Compare May 7, 2024 15:28

levythu mentioned this pull request May 14, 2024

Support multi-pass prefetch config in CacheParams, and shard estimator will correctly propagate its storage cost meta-pytorch/torchrec#2000

Closed

levythu force-pushed the export-D56908989 branch from 48773b2 to 25de7e0 Compare May 15, 2024 18:21

levythu force-pushed the export-D56908989 branch from 25de7e0 to e84feb7 Compare May 15, 2024 20:42

levythu force-pushed the export-D56908989 branch from e84feb7 to 47823df Compare May 15, 2024 20:43

levythu force-pushed the export-D56908989 branch from 47823df to 143e8f7 Compare May 15, 2024 21:55

levythu force-pushed the export-D56908989 branch from 143e8f7 to cf76b06 Compare May 16, 2024 02:53

levythu force-pushed the export-D56908989 branch from cf76b06 to 3e2f269 Compare May 16, 2024 02:54

levythu force-pushed the export-D56908989 branch from 3e2f269 to 3fa2a87 Compare May 16, 2024 14:25

levythu force-pushed the export-D56908989 branch from 8a1a461 to 0e08aa9 Compare May 16, 2024 15:02

levythu force-pushed the export-D56908989 branch from 0e08aa9 to 6985d2b Compare May 16, 2024 15:03

levythu force-pushed the export-D56908989 branch from 6985d2b to 1036df2 Compare May 16, 2024 15:13

levythu force-pushed the export-D56908989 branch from 1036df2 to 7bff08d Compare May 16, 2024 15:15

levythu force-pushed the export-D56908989 branch from 7bff08d to 86107ca Compare May 16, 2024 19:34

levythu force-pushed the export-D56908989 branch from 86107ca to 2b4eedf Compare May 17, 2024 00:03

levythu force-pushed the export-D56908989 branch from 2b4eedf to 3386341 Compare May 17, 2024 00:03

facebook-github-bot closed this in 578ab67 May 17, 2024

facebook-github-bot added the Merged label May 17, 2024

Uh oh!

Conversation

levythu commented May 7, 2024

Context

MultiPass Prefetch (MPP)

What's in the patch?

Uh oh!

facebook-github-bot commented May 7, 2024

Uh oh!

netlify Bot commented May 7, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

❌ Deploy Preview for pytorch-fbgemm-docs failed.

Uh oh!

facebook-github-bot commented May 7, 2024

Uh oh!

facebook-github-bot commented May 15, 2024

Uh oh!

facebook-github-bot commented May 15, 2024

Uh oh!

facebook-github-bot commented May 15, 2024

Uh oh!

facebook-github-bot commented May 15, 2024

Uh oh!

facebook-github-bot commented May 16, 2024

Uh oh!

facebook-github-bot commented May 16, 2024

Uh oh!

facebook-github-bot commented May 16, 2024

Uh oh!

facebook-github-bot commented May 16, 2024

Uh oh!

facebook-github-bot commented May 16, 2024

Uh oh!

facebook-github-bot commented May 16, 2024

Uh oh!

facebook-github-bot commented May 16, 2024

Uh oh!

facebook-github-bot commented May 17, 2024

Uh oh!

facebook-github-bot commented May 17, 2024

Uh oh!

facebook-github-bot commented May 17, 2024

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

netlify Bot commented May 7, 2024 •

edited

Loading