Support System Allocated Memory (SAM) #701
Draft
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
Signed-off-by: Rong Ou <[email protected]>
Signed-off-by: Rong Ou <[email protected]>
lijinf2
reviewed
Aug 5, 2024
| @@ -98,7 +98,10 @@ cat <<EOF | |||
| --spark_confs spark.python.worker.reuse=true \ | |||
| --spark_confs spark.master=local[$local_threads] \ | |||
| --spark_confs spark.driver.memory=128g \ | |||
| --spark_confs spark.rapids.ml.uvm.enabled=true | |||
| --spark_confs spark.rapids.ml.uvm.enabled=false \ | |||
| --spark_confs spark.rapids.ml.sam.enabled=true \ | |||
There was a problem hiding this comment.
Does sam work on non-GH machines? Will it fall back to uvm?
There was a problem hiding this comment.
Only when HMM is supported:
$ nvidia-smi -q | grep Addressing
Addressing Mode : HMM
There was a problem hiding this comment.
What will happen if HMM is not supported but we've enabled SAM?
There was a problem hiding this comment.
Invoking the RMM system mr would cause a crash. I guess we should figure out what value to default to that's the most convenient to us.
There was a problem hiding this comment.
Agree.
Does our ci machine support RMM? Nightly ci executes run_benchmark.sh with a A100 40G GPU I think.
There was a problem hiding this comment.
It should work as long as we install the open source driver. These are the requirements:
- NVIDIA CUDA 12.2 with the open-source r535_00 driver or newer.
- A sufficiently recent Linux kernel: 6.1.24+, 6.2.11+, or 6.3+.
- A GPU with one of the following supported architectures: NVIDIA Turing, NVIDIA Ampere, NVIDIA Ada Lovelace, NVIDIA Hopper, or newer.
- A 64-bit x86 CPU.
Signed-off-by: Rong Ou <[email protected]>
Signed-off-by: Rong Ou <[email protected]>
Signed-off-by: Rong Ou <[email protected]>
wbo4958
reviewed
Aug 5, 2024
docs/site/configuration.md
Outdated
| | spark.rapids.ml.uvm.enabled | false | if set to true, enables [CUDA unified virtual memory](https://developer.nvidia.com/blog/unified-memory-cuda-beginners/) (aka managed memory) during estimator.fit() operations to allow processing of larger datasets than would fit in GPU memory | | ||
| | spark.rapids.ml.sam.enabled | false | if set to true, enables System Allocated Memory (SAM) on [HMM](https://developer.nvidia.com/blog/simplifying-gpu-application-development-with-heterogeneous-memory-management/) or [ATS](https://developer.nvidia.com/blog/nvidia-grace-hopper-superchip-architecture-in-depth/) systems during estimator.fit() operations to allow processing of larger datasets than would fit in GPU memory | | ||
| | spark.rapids.ml.sam.headroom | None | when using System Allocated Memory (SAM) and GPU memory is oversubscribed, we may need to reserve some GPU memory as headroom to allow other CUDA calls to function without running out memory. Set a size appropriate for your application | | ||
| | spark.executorEnv.CUPY_ENABLE_SAM | 0 | if set to 1, enables System Allocated Memory (SAM) for CuPy operations. This enabled CuPy to work with SAM, and also avoid unnecessary memory coping | |
There was a problem hiding this comment.
Probably, do we need to add spark.driverEnv.CUPY_ENABLE_SAM when the spark is in local mode?
eordentlich
reviewed
Aug 6, 2024
docs/site/configuration.md
Outdated
| | spark.rapids.ml.uvm.enabled | false | if set to true, enables [CUDA unified virtual memory](https://developer.nvidia.com/blog/unified-memory-cuda-beginners/) (aka managed memory) during estimator.fit() operations to allow processing of larger datasets than would fit in GPU memory | | ||
| | spark.rapids.ml.sam.enabled | false | if set to true, enables System Allocated Memory (SAM) on [HMM](https://developer.nvidia.com/blog/simplifying-gpu-application-development-with-heterogeneous-memory-management/) or [ATS](https://developer.nvidia.com/blog/nvidia-grace-hopper-superchip-architecture-in-depth/) systems during estimator.fit() operations to allow processing of larger datasets than would fit in GPU memory | | ||
| | spark.rapids.ml.sam.headroom | None | when using System Allocated Memory (SAM) and GPU memory is oversubscribed, we may need to reserve some GPU memory as headroom to allow other CUDA calls to function without running out memory. Set a size appropriate for your application | | ||
| | spark.executorEnv.CUPY_ENABLE_SAM | 0 | if set to 1, enables System Allocated Memory (SAM) for CuPy operations. This enabled CuPy to work with SAM, and also avoid unnecessary memory coping | |
There was a problem hiding this comment.
enabled Cupy -> enables Cupy
coping -> copying
Signed-off-by: Rong Ou <[email protected]>
Signed-off-by: Rong Ou <[email protected]>
Signed-off-by: Rong Ou <[email protected]>
Signed-off-by: Rong Ou <[email protected]>
Signed-off-by: Rong Ou <[email protected]>
Signed-off-by: Rong Ou <[email protected]>
Signed-off-by: Rong Ou <[email protected]>
Signed-off-by: Rong Ou <[email protected]>
Signed-off-by: Rong Ou <[email protected]>
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.
Allowing running benchmarks using System Allocated Memory (SAM) on HMM and Grace Hopper systems. This depends on cupy/cupy#8442.