In hyper-scale bare-metal clouds, main system DRAM accounts for up to 50% of total server bill-of-materials cost, yet huge percentages of allocated physical memory remain cold. Linux Kernel Memory Tiering combined with Compute Express Link (CXL 2.0) pooled memory attaches disaggregated DRAM over PCIe Gen5 interfaces as CPUless NUMA nodes, allowing the kernel to automatically promote hot working sets to local DDR5 and demote cold pages to CXL expansion pools.
The Architecture of CXL NUMA Nodes & AutoNUMA Demotion
How the Linux page reclamation subsystem differentiates memory performance tiers:
CXL.mem expansion cards are enumerated by the Linux kernel as independent, CPUless NUMA nodes (e.g. Node 1 has memory but 0 cores). With sysctl vm.numa_zonelist_order and kernel memory tiering enabled (/sys/devices/virtual/memory_tiering/), kswapd reclaims memory by migrating inactive pages from Node 0 (Fast DRAM) to Node 1 (CXL DRAM) via DMA rather than swapping to NVMe disk, preserving microsecond access latency.
Memory Subsystems Compared
| Memory Tier | Physical Interface | Read Access Latency | Bandwidth Capacity |
|---|---|---|---|
| Tier 0: Direct Socket DDR5 | Direct CPU Memory Controller | 70 – 90 nanoseconds | 300 – 400 GB/s per socket |
| Tier 1: CXL 2.0 Expansion DRAM | PCIe 5.0 x16 CXL.mem Protocol | 170 – 220 nanoseconds | 64 GB/s per x16 lane |
| Tier 2: Enterprise NVMe zRAM/Swap | NVMe PCIe 4.0/5.0 SSD Block | 10 – 30 microseconds | 7 – 14 GB/s sequential |
Kernel Memory Tiering Sysfs Configuration
Enabling proactive page promotion and demotion across NUMA nodes:
# Enable kernel NUMA page demotion to CXL node
echo 1 > /proc/sys/kernel/numa_balancing
echo 1 > /proc/sys/kernel/numa_balancing_demote
# Configure AutoNUMA scan rate for hot page promotion
echo 1000 > /proc/sys/kernel/numa_balancing_scan_period_min_ms
echo 60000 > /proc/sys/kernel/numa_balancing_scan_period_max_ms
echo 256 > /proc/sys/kernel/numa_balancing_scan_size_mb
# Verify memory tiering node assignment
cat /sys/devices/virtual/memory_tiering/memory_tier*/nodelistExplore Enterprise Cloud & High-Throughput Infrastructure
Scale high-density bare-metal environments with zero latency compromises. Read our guide on Linux Kernel Memory Compaction & Proactive Reclaim, explore Node.js Unix domain socket IPC on WebDesigner.la SCM_RIGHTS Sockets, review vector symbolic ontologies on LinkDepot Hyperdimensional Computing, or consult with our cloud systems architects.