What is CXL (Compute Express Link) memory tiering and how does it transform Linux memory management?

Compute Express Link (CXL 2.0 / 3.0) is an open industry standard built on PCIe Gen 5 that enables CPUs to access external, byte-addressable memory devices (CXL.mem) with cache-coherent, load/store semantics at sub-microsecond latencies (typically 150–200 ns vs 80 ns for local DRAM). Linux kernel 5.18+ and 6.x introduced native Memory Tiering, which models CXL devices not as ordinary swap devices, but as a secondary, slower NUMA memory tier (Slow Tier). Active hot pages reside in expensive local DDR5 (Top Tier), while idle or cold pages are automatically migrated down to shared or pooled CXL memory, expanding server memory capacity by 4x to 8x at a fraction of the cost.

How does kswapd perform page demotion instead of disk swapping?

In traditional Linux memory management, when memory pressure reaches the watermark, the kernel page-stealing daemon (kswapd) reclaims memory by flushing dirty anonymous pages to disk swap or discarding file cache pages. Under tiered memory architectures, kswapd checks if a slower memory tier exists. If available, kswapd allocates pages in the CXL NUMA node and executes an in-kernel page migration (migrate_pages) to demote the cold page directly into CXL RAM. This avoids expensive NVMe/SATA I/O and context switches entirely, preserving application execution without catastrophic swap thrashing.

How does AutoNUMA promote hot pages back to fast DDR5 memory?

If an application thread frequently reads or writes a memory page that was previously demoted to CXL, the CPU triggers NUMA hinting page faults (unmapping the PTE or clearing the Present bit temporarily). The kernel's AutoNUMA subsystem intercepts the fault, tracks access frequency in page metadata, and if the access pattern qualifies as "hot", automatically migrates the page back up into local DDR5 memory (promotion). This bidirectional balancing ensures memory latency stays near native DDR5 speeds for active working sets.

What is vm.demote_scale_factor and how does it govern kernel reclamation behavior?

The sysctl parameter "vm.demote_scale_factor" (introduced in Linux 6.0) scales the memory pressure threshold at which kswapd begins proactively demoting pages from the top tier before reaching the emergency low watermark. By setting a higher scale factor (e.g. 200 or 300), system administrators ensure that local DRAM maintains a healthy cushion of free pages, preventing synchronous direct reclaim stalls during sudden allocation spikes.

Linux CXL Memory Tiering & Page Demotion Studio | Kernel Tiered NUMA Architecture

Verified Specification | Standardized Formulas | Instant Precision

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

Linux Kernel 6.x ABI CXL 2.0 / 3.0 Tiered NUMA Zero-Disk Dynamic Demotion

Linux CXL Memory Tiering & Page Demotion Studio

Model Linux kernel Compute Express Link (CXL) tiered memory architecture. Simulate kswapd cold page demotion, AutoNUMA hot page promotion, benchmark working-set latencies, and configure kernel sysctl tuning.

1. Hardware Memory Tier Topography

Top Tier (Fast Local DDR5 DRAM): Bottom Tier (CXL 2.0 Pooled Memory Node):

vm.demote_scale_factor:

AutoNUMA Promotion?

2. Application Workload & Memory Access Skew

Total Active Working Set: 320 GiB

Cold Data Fraction (% idle): 70%

DRAM Top Tier Load

96.0 GiB

75.0% utilized (32GB free)

CXL Bottom Tier Load

224.0 GiB

43.8% utilized (288GB free)

Effective Mean Latency

94.2 ns

vs 5,000,000 ns NVMe swap!

Demotion Throughput

14,200 p/s

55.4 MB/s zero-disk migration

Server RAM Cost Saved

58.4%

Pooled CXL TCO efficiency

3. Dynamic Page Demotion & Promotion Architecture

Hot Pages (DDR5 Tier 0) Demoted Cold Pages (CXL Tier 1) In-Flight Page Migration

kswapd demotes cold pages downwards; AutoNUMA promotes hot pages upwards

Linux CXL Memory Tiering & Page Demotion Studio

1. Hardware Memory Tier Topography

2. Application Workload & Memory Access Skew

3. Dynamic Page Demotion & Promotion Architecture

4. Linux Production Kernel sysfs & sysctl Configuration

🔍 Companion Workbenches & Verification Tools

SQL Query Formatter & Beautifier

Terraform & OpenTofu HCL Infrastructure Architect & Security Linter

Distributed Consensus, Raft & Paxos State Machine Studio

Linux io_uring & Asynchronous Zero-Copy I/O Architecture 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