Guest-Side Trick Cuts Virtual Machine Near-Memory Use by 50-70%

50-70% less precious near memory consumed for the same workload performance - that's what a new host-agnostic memory tiering technique delivers inside virtual machines.

The Huge-Page Blind Spot

Host-based memory tiering solutions scan the physical address space for hot huge pages and promote them to fast (near) memory. But in a virtualized environment, the host sees guest physical pages, not the actual access patterns inside the guest. When frequently accessed data are scattered across the guest physical address space - or when only a few subpages of a huge page are hot - the whole huge page looks hot to the host. Result: sparsely accessed huge pages get placed in expensive near memory, wasting capacity.

Consolidation Inside the Guest

The authors propose a technique that lives entirely inside the guest OS. It exploits the two-level address translation (guest physical to machine physical) to remap scattered and skewed accesses onto a small set of contiguous guest physical address ranges. By consolidating hot subpages into a handful of guest physical huge pages, those pages now appear densely hot to the host. The host's tiering policy then naturally promotes only the dense pages, leaving the sparse pages in far memory.

Crucially, the guest does not need any hypervisor cooperation. The technique is host-agnostic and works with any existing host-based tiering solution - it just changes the pattern the host observes.

Measured Gains: 50-70% Less Near Memory, 10-13% Faster at Scale

Evaluation on standalone real-world benchmarks (SPEC CPU, Redis, memcached) paired with a state-of-the-art host tiering scheme shows near-memory consumption drops 50-70% without degrading performance. At data-center scale, with multiple VMs contending for near memory, the consolidation frees enough bandwidth to improve overall workload throughput by 10-13% while keeping memory total cost of ownership flat.

This technique turns the guest from a passive consumer into an active participant in memory tiering - expect hypervisors and OS designers to start rethinking where the tiering intelligence lives.

Source: Efficient Memory Tiering in a Virtual Machine
Domain: arxiv.org