Red Hat engineers are looking at ways to make memory tiering work better for QEMU virtual machines, Phoronix reports, with an eye on using NVMe SSDs in more situations so VMs don’t have to sit entirely in expensive DRAM. Peter Xu presented the work this week at KVM Forum in Vancouver.
The options on the table include a new maccess() system call, letting QEMU own a slow tier in user space, or using existing kernel APIs to migrate hot pages, which currently sounds like the lead because it needs no kernel changes and only small QEMU ones. That path could get better still if the kernel’s MADV_PAGEOUT behavior is patched. This is a status talk, not a feature you can turn on in a release tomorrow.
My take
RAM prices have everybody thinking about slower tiers again. If QEMU can keep the hot pages in DRAM and spill the rest to NVMe without turning a VM into molasses, that’s real money for anyone running a lot of guests. I’ll wait for patches and numbers, not just a conference slide.
God Bless, and May the Source Be with You!