[QST]what is the meaning of RMM in configs.md

[YeahNew]

What is your question?
I'm sorry I don't understand the meaning of RMM in configs.md, it seems to be a resource? At the same time, I found that Rmm.class in cudf-0.14-cuda10-2.jar is the class used to initialize GPU memory. Is there any connection between them?
When I set spark.rapids.memory.gpu.pool.enabled=true, the query#5 was executed for 1.2mins, when it was false, query#5 was executed for 1.6mins. This setting approximately accelerated the query by 33.3%. I want to know the internal reasons.
I also don't quite understand the explanation of this parameter in configs.md.
Can you help me?

[jlowe]

I don't understand the meaning of RMM

RMM stands for RAPIDS Memory Manager. It's designed to accelerate CUDA memory allocations. Normally CUDA memory operations like cudaMalloc and cudaFree are very expensive because they need to update page tables on the device. RMM allocates memory in large chunks and retains those allocations, turning them into a pool of memory that it then sub-allocates to satisfy memory requests. This is much, much faster in practice than calling cudaMalloc and cudaFree directly.

I also don't quite understand the explanation of this parameter in configs.md.

Note that this is covered in our tuning guide along with other important settings for performance such as using pinned host memory. If the explanation is still not clear after reading the tuning guide, please let us know and we'll work to improve the documentation.

[YeahNew]

Thank you, I have got it