13–15 Nov 2023
America/New_York timezone

A move_pages() equivalent for physical memory

13 Nov 2023, 17:35
25m
"Potomac G" (Omni Richmond Hotel)

"Potomac G"

Omni Richmond Hotel

80
Compute Express Link MC Compute Express Link MC

Speakers

Mr Gregory Price (MemVerge Inc) Svetly Todorov (MemVerge)

Description

CXL-aware job schedulers, memory managers, and userspace tiering solutions depend on page migration syscalls to reallocate resources across nodes. Currently, these calls enable movement of memory associated with a specific PID. Moves can be requested in coarse, process-sized strokes (as with migrate_pages), and on specific virtual pages (via move_pages). However, a number of profiling mechanisms provide information at a systemwide granularity: the IDLE bit is cleared on reads/writes of physical pages, /proc/zoneinfo breaks PFN-space into NUMA nodes, and PEBS/IBS can be configured to track frequency of accesses to physical addresses. Information from these sources facilitates systemwide resource management, but in order to actually perform said management, their outputs must be converted back to virtual addresses and re-associated with specific PIDs. Doing this reverse-translation outside of the kernel requires considerable space and compute. It can potentially be avoided if the contents of physical memory can be migrated directly.
    We propose a syscall, analogous to move_pages, that migrates the contents of physical pages between NUMA nodes. We present our investigations as to how this call might be integrated with minimal changes to the mm/migrate.c codebase. We also explore a prototype tiering framework that could make use of this call, and its limitations.

Primary authors

Mr Gregory Price (MemVerge Inc) Svetly Todorov (MemVerge)

Presentation materials

Diamond Sponsors
Platinum Sponsor
Gold Sponsors
Silver Sponsors
Catchbox Sponsor
Livestream Sponsors
T-Shirt Sponsor
Conference Services Provided by