As the programming models for GPU-based high-performance computing applications are evolving, HMM is helping us integrate the GPU memory management more closely with the kernel's virtual memory management. As a result we can provide a shared virtual address space with demand-paging and page-based migrations of anonymous pages to/from device memory. A patch series by AMD [1, 2] to add support for cache-coherent, CPU-accessible device memory has brought up some fairly fundamental questions about HMM and its interaction with virtual memory management, page cache and file systems. We'd like to use the chance of getting together for a BoF [3] at LPC to raise awareness for HMM outside the GPU driver code, identify gaps in the architectural documentation and clarify our priorities for future development.
Thank you, Daniel, for suggesting the BoF and getting it scheduled. It's set for Friday, 10am Pacific, 1pm Eastern, 5pm UTC.
I am registered at LPC. Daniel got a speaker's pass. We're still trying to work something out for Alex.
I hope to see many of you on Friday.
Best regards, Felix
[1] https://patchwork.freedesktop.org/series/94611/ [2] https://patchwork.freedesktop.org/series/90706/ [3] https://linuxplumbersconf.org/event/11/contributions/1123/