Avoid unnecessary calls to cuFuncSetAttribute #2678
+3
−1
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
Calling cuFuncSetAttribute to set CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES is required to launch kernels using more than 48 kB of local memory[1] (CUDA dynamic shared memory). Without this, cuLaunchKernel fails with CUDA_ERROR_INVALID_VALUE. However, calling cuFuncSetAttribute introduces synchronisation in the CUDA runtime which blocks its execution until all H2D/D2H memory copies are finished (don't know why), therefore effectively blocking kernel launches from overlapping with memory copies. This introduces significant performance degradation in some workflows, specifically in applications launching overlapping memory copies and kernels from multiple host threads into multiple CUDA streams to the same GPU. Avoid the CUDA runtime synchronisation causing poor performance by removing the cuFuncSetAttribute call unless it's strictly necessary. Call it only when a specific carveout is requested by user (using env variables) or when the kernel launch would fail without it (local memory size >48kB). Good performance is recovered for default settings with kernels using little or no local memory. No performance effects were observed for kernel execution time after removing the attribute across a wide range of tested kernels using various amounts of local memory. [1] Related to the 48 kB static shared memory limit, see the footnote for "Maximum amount of shared memory per thread block" in https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#features-and-technical-specifications-technical-specifications-per-compute-capability
|
This issue came from a customer benchmark. I cannot share the code, but I tested the change after introducing a configurable amount of shared memory in their benchmark and observed the following results: Also, for the record, here's how the CUDA runtime synchronisation looks like in nsys: without cuFuncSetAttribute (overlapping operations) with cuFuncSetAttribute (locking occurs): |
|
intel/llvm PR: intel/llvm#16928 |
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.
Calling
cuFuncSetAttributeto setCU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTESis required to launch kernels using more than 48 kB of local memory[1] (CUDA dynamic shared memory). Without this,cuLaunchKernelfails withCUDA_ERROR_INVALID_VALUE. However, callingcuFuncSetAttributeintroduces synchronisation in the CUDA runtime which blocks its execution until all H2D/D2H memory copies are finished (don't know why), therefore effectively blocking kernel launches from overlapping with memory copies. This introduces significant performance degradation in some workflows, specifically in applications launching overlapping memory copies and kernels from multiple host threads into multiple CUDA streams to the same GPU.Avoid the CUDA runtime synchronisation causing poor performance by removing the
cuFuncSetAttributecall unless it's strictly necessary. Call it only when a specific carveout is requested by user (using env variables) or when the kernel launch would fail without it (local memory size >48kB). Good performance is recovered for default settings with kernels using little or no local memory.No performance effects were observed for kernel execution time after removing the attribute across a wide range of tested kernels using various amounts of local memory.
[1] Related to the 48 kB static shared memory limit, see the footnote for "Maximum amount of shared memory per thread block" in https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#features-and-technical-specifications-technical-specifications-per-compute-capability