Open abellina opened 2 years ago
One issue that was raised was that a DeviceMemoryBuffer
could instead be composed using a java "MemoryResource", rather than having to create other classes like CudaMemoryBuffer
to accomplish this. We could then pass other resources like a potential "AsyncMemoryResource" instance, or some hierarchy of other resources (the logging resource + pool for example). Ideally this issue would tackle this problem as well.
This issue has been labeled inactive-30d
due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d
if there is no activity in the next 60 days.
This issue has been labeled inactive-90d
due to no recent activity in the past 90 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.
We currently support in the Java side some combinations of allocators and wrappers that make it easy to setup pools (default, arena, and now async), backing allocators (cuda, managed memory). But it is getting to the point that the options in these allocators don't all fit the min/max pool size pattern.
For general users of the cuDF JNI side of things the above combinations may not be desired. We propose exposing an RMM that would allow better composability of the various allocators we use and test with (note this is a good chance to clean up some tech debt as well for allocators we no longer use or intend to support).