KVM Approximate-LRU Page Eviction

Replaced the FIFO page-eviction policy in the Linux KVM hypervisor’s shadow-paging MMU with an approximate LRU (second-chance clock) policy, so that frequently accessed guest pages stay resident and page faults drop.

Working in the kernel’s mmu.c, we added lightweight per-page recency tracking and reworked mmu_zap_oldest_pages() to evict by access recency rather than insertion order, keeping the eviction path thread-safe across multiple vCPUs. We benchmarked against FIFO with stress-ng, redis-benchmark, and flamegraph profiling on an AWS bare-metal instance: Approximate LRU reduced page faults (112,016 vs. 117,375 under stress-ng) and sustained lower latency at high load, at the cost of slightly higher CPU overhead from the recency tracking.

Tech stack: C, Linux kernel, KVM, QEMU.