[FEA] Update CUDF host memory allocation APIs so allocation is pluggable. #9862
Assignees
Labels
reliability
Features to improve reliability or bugs that severly impact the reliability of the plugin
task
Work required that improves the product but is not user facing
Comments
This was referenced Nov 27, 2023
3 tasks
copy-pr-bot bot
pushed a commit
to rapidsai/cudf
that referenced
this issue
Nov 28, 2023
spark-rapids requires the ability to set DefaultHostMemoryAllocator Contributes to NVIDIA/spark-rapids#9862 Signed-off-by: Gera Shegalov <gera@apache.org>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Is your feature request related to a problem? Please describe.
rapidsai/cudf#13924 added in a HostMemoryAllocator interface. But it also made using a non-default one on a per-API basis. This is what we wanted when I thought we could get memory limits without using the retry framework. After hitting some design issues it became clear that we have to use the retry framework for this. With that in mind we should just update the code so we have an API to set the default allocator use and that all allocations from java use that default allocator. We can either deprecate the per-API allocators or just remove them because I don't think that they are being used any more.
On the plugin side we will want to update the initialization code for HostAlloc to set this if there is a memory limit.
At the same time we can probably remove the reservation API from the pinned memory pool. It is not needed with the new design.
The text was updated successfully, but these errors were encountered: