11–13 Dec 2025
Asia/Tokyo timezone

Mempolicy is dead, long live memory policy!

Not scheduled
20m
Kernel Memory Management MC Kernel Memory Management MC

Speaker

Gregory Price (Meta)

Description

The default global mempolicy is inclusive of all NUMA nodes - where the fallback allocation behavior is typically defined by NUMA distances. Tasks and cgroups are then expected to opt-in to more restrictive policies via set_mempolicy and cpusets interfaces.

This is the opposite of a typical isolation mechanisms - and leads to global resource (such as unmapped pagecache) having poor isolation controls when accessed by tasks with different memory policies.

What if, instead, the default policy was restrictive (DRAM/Default-tier only), and tasks/cgroups were required to opt-in to tiered memory (via a set_mempolicy, cgroup/cpusets, or even new per-file/filesystem policies). We will discuss one design and the implications for various subsystems (reclaim, compaction, migration, pagecache vs anon allocations, cgroups/cpusets controls, etc).

Primary author

Presentation materials

There are no materials yet.