What is Non-Uniform Memory Access (NUMA) in modern multi-socket servers?

In modern multi-socket server hardware (such as dual AMD EPYC or Intel Xeon systems), physical RAM is divided across NUMA nodes attached directly to specific CPU sockets. Accessing local RAM directly wired to a CPU core occurs over ultra-fast, wide memory channels (~60-70 ns latency). Accessing remote RAM attached to another socket requires traversing inter-socket interconnects (Intel UPI or AMD Infinity Fabric), adding 60-100% latency penalty (~110-140 ns) and consuming inter-socket interconnect bandwidth.

What is the difference between MPOL_BIND, MPOL_INTERLEAVE, and MPOL_PREFERRED?

Linux provides thread and process memory allocation policies via set_mempolicy() and mbind(). MPOL_BIND strictly restricts allocations to specified NUMA nodes, failing with ENOMEM if those nodes exhaust memory. MPOL_PREFERRED attempts allocation on the designated node first, falling back to other nodes under memory pressure. MPOL_INTERLEAVE round-robins page allocations page-by-page (4 KiB interleaved) across specified nodes, maximizing memory bandwidth across all memory controllers for cache-agnostic workloads (e.g. databases, in-memory caches).

How does Linux Automatic NUMA Balancing (AutoNUMA) detect and migrate remote pages?

When kernel.numa_balancing = 1 is enabled, the kernel periodically marks memory pages as unmapped or inaccessible in the page table (PROT_NONE / PTE numa hint). When a thread on a CPU accesses this page, a NUMA hinting page fault occurs. The kernel records which CPU/node triggered the fault. If repeated accesses originate from a remote node, the kernel transparently calls migrate_pages() to move the physical page to the accessing node local RAM, eliminating ongoing cross-socket latency.

Why does AutoNUMA sometimes harm high-performance database workloads like PostgreSQL and Redis?

AutoNUMA incurs CPU overhead from periodic scan passes and hinting page faults. For workloads with shared read-heavy data accessed equally by threads on multiple sockets (like database buffer pools), pages bounce back and forth between sockets in an expensive "NUMA thrashing" loop. Production database administrators often disable kernel.numa_balancing and use numactl --interleave=all to evenly distribute buffer pool allocations across all nodes.

Linux NUMA Memory Balancing & Interleave Policy Architecture Studio

Verified Specification | Standardized Formulas | Instant Precision

Secure & Private (Zero Data Retention) Free Access • No Sign-Up

Linux NUMA Memory Balancing & Interleave Policy Studio

Multi-socket kernel memory simulator modeling NUMA node topologies, interconnect fabric latency penalties, numactl allocation policies, and AutoNUMA page migration.

Dual-Socket / UPI / IF

Memory Policy (numactl)

Active Workload Node Affinity

Automatic NUMA Balancing

Working Set Memory Size (GB)

16 GB 64 GB 256 GB

Avg Memory Access Latency

78.5 ns

+18.5 ns cross-socket penalty

Remote Memory Access Ratio

25.0 %

16.0 GB accessed across interconnect

Aggregate Memory Bandwidth

384 GB/s

Dual memory controllers saturated

AutoNUMA Page Migrations

Active

migrate_pages() scanning hints

Dual-Socket NUMA Topology & Interconnect Traffic ● Interconnect: 64 GB/s UPI Bus (Normal)

Linux Production numactl Command & Kernel sysctl Config

Linux NUMA Memory Balancing & Interleave Policy Studio

🔍 Companion Workbenches & Verification Tools

Linux eBPF bpf_iter & Kernel Task Iterators Studio | In-Kernel Streaming Observability

Memory-Mapped Files (mmap), Page Cache & Linux Virtual Memory Studio

Linux pidfd Process Lifecycle & Race-Free Tracking Studio

Linux Transparent Huge Pages (THP) & Khugepaged Studio

Why Smart Users Are Switching to This Free Toolkit

The Browser Feature Most People Never Use

Stop Paying for Software You Can Get Free

File Conversion Tricks That Save Hours Every Week

This Free Tool Replaced 5 Paid Apps for Our Team

What Your Browser Can Do Without Any Extensions