Research describes fault isolation in a multi-process GPU service and VMM-based active-standby recovery; the authors report zero runtime overhead for the isolation mechanism.
Published on May 27, 2026, the report describes a multi-process GPU service combining driver-level isolation with VMM-based active-standby recovery. The approach aims to confine reachable MMU faults to the affected client and enable service recovery. Its scope is fault isolation and recovery in multi-process GPU workloads.
The researchers report zero runtime overhead for the isolation mechanism. The available material does not detail the evaluation methodology or the limits of this result. To verify the claim, consult the original publication and examine the test conditions, definition of overhead, and fault scenarios evaluated.