Go Huge or Go Home: Flexible Hugepage Swapping for Cloud VMs

BBB-Link: https://bbb.tu-dresden.de/b/mat-xin-oyh-xzn

Presentation Language: English

Memory overcommitment in cloud data centers is a critical challenge: VMs are routinely allocated more memory than they actually use, yet reclaiming that cold memory efficiently remains an unsolved problem. Existing approaches rely on general-purpose OS swap mechanisms that were never designed with virtualized workloads in mind, leaving significant performance and utilization gains on the table. In this talk, I present a userspace framework built specifically for VMs running on Linux/KVM. It exposes a flexible API for defining custom memory reclamation policies and introduces huge page-based swapping to reduce virtualization overhead. It enables zero-copy I/O over shared VM memory and dynamically balances memory pressure across co-located VMs. I also demonstrate its flexibility by integrating it with SPDK, showcasing support for advanced storage configurations such as software RAID. This talk will walk through the design, key implementation challenges, and lessons learned from deploying a userspace approach to a problem traditionally handled deep inside the kernel.