Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore a fundamental redesign of Linux memory policy architecture in this 14-minute conference talk that challenges the current inclusive NUMA node allocation model. Learn about the limitations of the existing mempolicy system where the default global policy includes all NUMA nodes with fallback behavior defined by NUMA distances, requiring tasks and cgroups to opt into restrictive policies through set_mempolicy and cpusets interfaces. Discover how this approach creates isolation problems, particularly with global resources like unmapped pagecache when accessed by tasks with different memory policies. Examine a proposed alternative design where the default policy becomes restrictive (DRAM/Default-tier only), requiring explicit opt-in for tiered memory access through various mechanisms including set_mempolicy, cgroup/cpusets, or new per-file/filesystem policies. Understand the wide-ranging implications this architectural change would have across multiple Linux subsystems including memory reclaim, compaction, migration, pagecache versus anonymous allocations, and cgroups/cpusets controls.
