The State Of Heterogeneous Memory Management At The End Of 2018
Heterogeneous Memory Management is the effort going on for more than four years that was finally merged to the mainline Linux kernel last year but is still working on adding additional features and improvements. HMM is what allows for allowing the mirroring of process address spaces, system memory to be transparently used by any device process, and other functionality for GPU computing as well as other device/driver purposes.
Jerome Glisse at Red Hat who has spearheaded Heterogeneous Memory Management from the start presented at last week's Linux Plumbers Conference on this unified memory solution.
Some of the future work still for HMM includes a new API for memory placement, new ways to expose the system topology to user-space, allowing different devices to access other device memory, and new automatic memory placement that can handle device memory. Glisse ended his presentation with, "We are blurring the boundary between CPU and device computation!"
Those wanting to learn more about the current state and future of Linux HMM can checkout Jerome's slide deck (PDF).
Jerome Glisse at Red Hat who has spearheaded Heterogeneous Memory Management from the start presented at last week's Linux Plumbers Conference on this unified memory solution.
Some of the future work still for HMM includes a new API for memory placement, new ways to expose the system topology to user-space, allowing different devices to access other device memory, and new automatic memory placement that can handle device memory. Glisse ended his presentation with, "We are blurring the boundary between CPU and device computation!"
Those wanting to learn more about the current state and future of Linux HMM can checkout Jerome's slide deck (PDF).
5 Comments