Deep Dive: NUMA Architecture and Memory Performance for VMs

[acmenezes]

Feature Type

Deep dive expanding existing tutorial

Description

Create a deep-dive companion to the existing numa-hugepages.adoc (782 lines, ~2,830 words) that covers advanced NUMA topology concepts, hugepage management strategies, and memory performance measurement. The current tutorial enables NUMA passthrough and hugepages but does not address multi-NUMA VMs, memory bandwidth measurement, or the interaction between NUMA balancing and pinned VMs.

Key Topics to Cover

• Multi-socket NUMA topology: Understanding 2-socket and 4-socket NUMA topologies, NUMA distances, and how they impact VM performance on different hardware
• NUMA distance measurement: Using numactl --hardware and lscpu --extended to understand the physical topology, interpreting NUMA distance tables
• Multi-NUMA node VMs: Configuring VMs that span multiple NUMA nodes -- when it makes sense (very large VMs) vs when to avoid it (latency-sensitive workloads)
• Hugepage pre-allocation strategies: Boot-time allocation (guaranteed, requires reboot) vs runtime allocation (flexible, may fail due to fragmentation) -- deciding which approach for production
• 2Mi vs 1Gi hugepage trade-offs: 2Mi is flexible and wastes less memory on small allocations; 1Gi reduces page table overhead to near zero but wastes memory in 1 GB increments -- decision criteria
• NUMA balancing (autonuma): How automatic NUMA balancing interacts with pinned VMs -- why you should disable it for dedicated-CPU VMs (prevents kernel from moving memory between nodes)
• Memory bandwidth measurement: Using Intel MLC (Memory Latency Checker) and STREAM benchmark inside VMs to validate NUMA-local vs cross-NUMA memory access performance
• NUMA-aware application deployment inside VMs: Configuring applications (databases, JVM) to be NUMA-aware within the guest OS
• Large VM (256 GB+) considerations: Page table memory overhead, transparent hugepage (THP) interaction, memory hotplug for dynamic resizing
• Troubleshooting memory performance: Identifying cross-NUMA access patterns, diagnosing TLB pressure, perf tools for memory profiling inside VMs

Value Proposition

The current tutorial enables the features. This deep dive helps users understand the underlying hardware architecture and make informed decisions about hugepage sizing, NUMA topology exposure, and memory performance validation. This is essential for workloads where memory access latency directly impacts application performance (databases, caches, HPC).

Suggested Module

modules/performance/