Linux kernel 6.18 shipped a per-CPU slab allocation caching layer called “sheaves” in late 2026, and Google’s Multi-Gen LRU (MGLRU) framework is now fully integrated across mainline kernels. With Linux holding 44.25% server OS market share in 2026 and 92% of virtual machines across AWS, Azure, and Google Cloud running on Linux, these memory management gains have measurable cost implications at fleet scale. Here’s how Linux, Windows, and macOS compare on memory efficiency in 2026, backed by kernel benchmarks and real-world telemetry.
Linux Memory Management Key Statistics (2026)
- MGLRU reduced kswapd CPU usage by 40% across Google’s Chrome OS fleet (Phoronix).
- Low-memory kills dropped 85% at the 75th percentile on Chrome OS devices running MGLRU (Phoronix).
- A clean Windows 11 install uses 2.2–3.0 GB of RAM at idle. Ubuntu Linux uses roughly 1 GB (Abanoub Hanna).
- Headless Linux servers operate at 250–400 MB idle memory, compared to 4–8 GB on Windows with SysMain active (Server Academy, Digital Trends).
- Linux kernel 6.18 introduced per-CPU “sheaves” for slab allocations, reducing lock contention on 64+ core servers (Linux Journal).
Linux vs Windows vs macOS: Idle Memory Footprint Comparison
How much RAM does each OS consume before any user application runs? The gap is substantial. A fresh Windows 11 install draws 2.2–3.0 GB at idle. With SysMain (formerly Superfetch) active, that number routinely reaches 4–5 GB on 16 GB systems. Pop! OS, a mainstream Ubuntu-based distribution, uses about 1 GB after a fresh boot on x64 hardware.
Headless Linux servers without a graphical environment sit at 250–400 MB. macOS lands between the two, though Apple’s memory compression (roughly 2.3x ratio) makes direct comparison tricky—much of what Activity Monitor reports as “memory used” is compressed data, not fully occupied physical address space. The practical result: Linux leaves more physical RAM free for actual workloads.
| Operating System | Idle RAM Usage | Notes |
|---|---|---|
| Windows 11 (clean install) | 2.2–3.0 GB | Baseline with updates, no third-party apps |
| Windows 11 (with SysMain) | 4–8 GB | Caching scales with total RAM |
| Ubuntu Linux (GNOME desktop) | ~1.0 GB | Pop! OS benchmark, fresh boot |
| Ubuntu Linux (headless server) | ~250–400 MB | No GUI, minimal services |
| macOS (Apple Silicon, 8 GB) | ~3.5–4.5 GB | Includes compressed and wired memory |
Linux MGLRU: Memory Reclamation Benchmarks Under Load
Multi-Gen LRU is the most consequential Linux memory management upgrade of recent years. Developed by Google engineer Yu Zhao and merged into the mainline kernel in 2022, MGLRU replaces the old two-list (active/inactive) page reclamation model with a multi-generational approach that tracks page usage at finer granularity.
Google tested MGLRU on tens of millions of Chrome OS devices and about one million Android devices before upstream integration. The fleet-wide numbers are hard to argue with: 40% less CPU spent on kswapd (the kernel daemon that reclaims memory pages), 85% fewer low-memory kills at the 75th percentile, and 18% faster app launch times at the 50th percentile on Android.
An independent lab evaluation published 960 data points across Apache Hadoop, Memcached, MongoDB, PostgreSQL, and Redis over 500+ hours of benchmark time. With 95% confidence intervals, all tested applications performed better under MGLRU for at least part of their benchmark matrices. Linus Torvalds personally endorsed the patch series.
Linux Memory Pressure Handling vs Windows and macOS
Each OS takes a different approach when physical RAM fills up. Linux offers the most granular control. The vm.swappiness parameter (adjustable from 0 to 200 since kernel 5.8) lets administrators control how aggressively the kernel reclaims anonymous pages versus file-backed pages. Paired with zswap—a compressed write-back cache—Linux can compress and store pages in RAM, deferring or skipping disk I/O entirely.
macOS runs its memory compressor automatically at about 2.3x compression ratios. It has no OOM Killer equivalent; it compresses first, then falls back to disk-based swap. This works well for desktop responsiveness but provides less control for server workloads. Windows starts writing to its pagefile from boot regardless of actual memory pressure, and SysMain pre-loads frequently used applications into RAM. That’s why a Windows system with 32 GB of RAM often reports 8–10 GB used at idle.
Linux Kernel 6.18 Sheaves: Slab Allocation Improvements (2026)
Released November 30, 2026, Linux kernel 6.18 introduced “sheaves”—a per-CPU caching layer for slab allocations. On high-core-count servers (64+ cores), the global slab allocator lock becomes a bottleneck during workloads with frequent small allocations and deallocations. Sheaves cache allocations at the per-CPU level, cutting lock contention and speeding up both allocation and freeing operations.
Concurrent with sheaves, kernel 6.18 also improved overall swapping performance under heavy workloads. For comparison, Windows Server 2026 does not expose an equivalent tuning mechanism. Its allocator improvements are bundled into opaque system updates, and administrators have no way to configure allocation caching strategy at this level. This matters for anyone monitoring RAM usage on Linux servers running memory-heavy applications.
Linux Memory Management Kernel 6.18 Feature Summary
| Feature | Linux 6.18 | Previous Kernels |
|---|---|---|
| Slab allocation caching | Per-CPU “sheaves” layer | Global slab allocator |
| Memory tiering | DRAM over NVMe-backed memory | Basic NUMA support |
| Swap subsystem | Improved under pressure | Standard swap paths |
Electron App RAM Usage on Windows vs Linux
Electron-based applications—each bundling a Chromium instance—are a growing memory problem in 2026, and they hit Windows harder than Linux. Discord’s Windows client can reach 4 GB of RAM under real-world conditions, far exceeding the “under 1 GB” baseline the company claims. Discord’s response was an auto-restart experiment that silently restarts the app after 30 minutes of idle time when memory exceeds 4 GB.
Microsoft Teams routinely sits at about 1 GB idle on Windows despite its migration from Electron to WebView2. On Linux, these same apps tend to consume less memory because the OS leaves more RAM available at baseline, and MGLRU reclaims pages from idle processes more aggressively. The absence of SysMain-style pre-caching on Linux also means less baseline memory contention for Electron apps to compete against.
Why Linux Memory Management Efficiency Matters at Scale
With Linux powering the majority of web servers and 92% of VMs across major cloud providers, even small memory efficiency gains compound into significant cost savings. A 1% improvement in memory reclamation across a fleet of tens of thousands of servers translates to millions of dollars in reduced hardware provisioning.
The growing Linux adoption rate across countries amplifies this effect. Combined with MGLRU’s 40% kswapd CPU reduction and 85% drop in low-memory kills, Linux’s memory stack—including configurable swappiness, zswap, and the new sheaves allocator—gives it a measurable edge in server and high-load scenarios. macOS remains competitive for desktop use where compression and App Nap keep responsiveness high. Windows trades baseline efficiency for aggressive caching through SysMain, an approach that works on systems with plenty of RAM but penalizes memory-constrained environments.
FAQs
How much RAM does Linux use at idle compared to Windows?
A clean Ubuntu desktop uses about 1 GB at idle. A headless Linux server uses 250–400 MB. Windows 11 uses 2.2–3.0 GB at idle and 4–8 GB with SysMain caching active.
What is MGLRU in the Linux kernel?
Multi-Gen LRU is a page reclamation framework developed by Google. It replaced the traditional two-list model with multi-generational tracking, reducing kswapd CPU usage by 40% and low-memory kills by 85% across Chrome OS devices.
Does macOS handle memory better than Linux?
macOS uses built-in memory compression at roughly 2.3x ratios, which keeps desktops responsive on limited RAM. Linux offers more control through configurable swappiness, zswap, and MGLRU, making it better suited for server workloads.
What are “sheaves” in Linux kernel 6.18?
Sheaves are a per-CPU caching layer for slab memory allocations introduced in kernel 6.18 (November 2026). They reduce lock contention on high-core-count servers during frequent small allocation and deallocation operations.
Why do Electron apps use more RAM on Windows?
Windows uses more baseline memory through SysMain pre-caching, leaving less free RAM. Linux’s MGLRU reclaims idle process pages more aggressively, and the lower baseline footprint gives Electron apps more room to operate.
