You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Restore: read the state as fast as the disk, or only what the guest touches #276
A checkpoint restored to the first login is ~160 ms once its state is in memory. From the runner's gp3 volume it is ~840 ms, of which ~680 is reading the ~163 MiB state (#275, labs 44542990433 and 44542999076). QEMU reads it ~250 ms slower than one sequential pass of the same file does. spin's hosts read from local NVMe, so the disk share there is smaller, but neither the read nor the copy into guest memory is free. Experiments, each a row of boot:restore-first, measured on the bench once it has local NVMe (spin-stack/ci#21):
Read the state the way the disk can.posix_fadvise(POSIX_FADV_SEQUENTIAL) and a large readahead on the file restore hands QEMU, or QEMU's file migration reading in larger blocks. Up to ~250 ms on gp3; on NVMe, whatever is left of it. Fingerprint unchanged.
Read only what the guest touches (restore on demand). Guest RAM backed by a file (memory-backend-file, share=on), saved with x-ignore-shared: the checkpoint holds only device state, and the restore maps the RAM file and faults pages in as the guest touches them. The first login touches a fraction of 163 MiB. To find out:
where the RAM file lives while the workspace is suspended (tmpfs keeps host memory, a disk takes the guest's writes);
how it composes with virtio-mem and the balloon's free page reporting;
it is a change to Spec.Contract, so the fingerprint moves (CLAUDE.md: allowed when it serves the goal; the release notes say so).
A smaller state. Most of 163 MiB in a 2 GiB guest is likely its page cache. Dropping clean caches, or reporting free pages, right before the save shrinks what is read. The save is spin's to trigger, so this is a contract with spin: what the guest is asked to do before a checkpoint.
Already rejected: mapped-ram with multifd (#246, 385 vs 386 ms, measured with the disk share unseparated; worth one row again once 1 says what the read costs).
Activity