GMEM: Generalized OS Memory Management for Accelerators – Weixi Zhu, Huawei
Domain-specific accelerators support virtual memory for programmability and resource isolation. Currently, their drivers keep reinventing the wheel, though the OS already supports memory management. This is because Linux’s memory management (MM) cannot be easily utilized for accelerators. So, the amount of memory management code keeps growing, introducing code bloating and extra bugs. This talk presents GMEM (Generalized Memory Management) which makes an invasive design change to Linux MM. GMEM decouples CPU-related implementations from hardware-independent memory management mechanisms, so drivers can register their hardware-specific functions and let the OS manage memory. Furthermore, GMEM provides better programmability for accelerators — encapsulating accelerators and CPU together within a single address space. This saves programmers from manually managing memory between two parallel address spaces. Furthermore, programmers can utilize CPU DRAM as a victim cache for accelerators, without adding a single line of code. GMEM is implemented on a Linux-based OS, openEuler. In this talk, Weixi will introduce GMEM’s design, discuss how drivers of commercial accelerators integrate with GMEM and demonstrate a case study of an existing commercial accelerator.
