The problem
What does the host scheduler know about the work represented by a vCPU or a device-feeder thread?
What current research shows
A host schedules threads and virtual CPUs as host-visible execution units. The guest or accelerator may have its own queues, dependencies, and timing. That creates a boundary where a locally reasonable host delay can affect work the host can’t see directly.
Where the evidence stops
The size of the hidden state depends on the hypervisor, device interface, guest scheduler, and telemetry available. This study doesn’t claim that one universal host policy solves the boundary.
What Valen Systems is testing
Record guest runnable state, device queue depth, CPU feeder delay, accelerator idle time, and end-to-end work completion under matched host policies.