KVM Forum 2021

Linux is a great hypervisor platform, but it does not yet provide a complete solution for overcommitting VM memory efficiently. We present lessons learned from implementing a self-adapting memory-overcommit solution based on QEMU/kvm and libvirt. We discuss the drawbacks of different memory backends, such as hugetlbfs and memfd, the challenges of coordinating memory ballooning and hypervisor swap via cgroups, and how out-of-the-box live-migration performance can be catastrophic without careful per-page reclamation hints. In addition, we describe how to collect VM memory stats to estimate memory needs, especially in the absence of metrics provided by balloon drivers. This includes leveraging idle page tracking to obtain insights into VM memory usage, using sampling to avoid excessive CPU overhead. Finally, we explain how to tie those input metrics and output knobs together to produce a complete solution for adaptively overcommitting VM memory efficiently in production systems.