GPU Computing · Cluster
GPU Virtualization
Partitioning, passthrough, and multi-tenancy.
- 01 The Interception Surface of a Virtual GPU A GPU presents a register file, a bus-mastering DMA engine, and an interrupt source at once; sharing one means deciding which of the three the host keeps control of.
- 02 PCI Passthrough and the Boundaries an IOMMU Actually Draws Why handing a GPU to a guest requires an isolation unit that no sibling device can bypass, and what changes when the device caches its own translations.
- 03 SR-IOV Virtual Functions and Static Partitioning of GPU Engines A single PCI function becomes many by replicating the parts that carry identity, and the GPU functions that cannot be replicated become the hard limits on how finely a card can be split.
- 04 Paravirtual GPUs and the Mediated Device Lifecycle How a host-side scheduler multiplexes one physical GPU across guests without giving any of them direct access, and what the create-configure-assign-teardown sequence has to guarantee.
- 05 Fixed-Slice Partitioning of Shared Engines Inside One Device Splitting a GPU into isolated instances without separate PCI functions requires slicing the caches and memory controllers, and the slices that cannot be cut set the floor on granularity.
- 06 Time-Slicing and What a GPU Context Switch Costs Preemption on a GPU is not one operation but several, each at a different granularity, and the cost of switching decides whether oversubscription is worth it.
- 07 Device Memory Oversubscription and Page Migration Selling more device memory than exists requires evicting pages to host memory, and the cost of that depends entirely on how the device detects and services a miss.
- 08 Tenant Isolation and the Channels Performance Interference Travels Through Separating address spaces does not separate throughput; the shared resources that remain determine how much one tenant can slow another and which controls can bound it.
- 09 Address Translation Across Host and Device A GPU access may traverse three translation stages, and where the stages live determines whether a shared address space is cheap, whether faults are recoverable, and what a migration can move.
- 10 Live Migration of a Guest That Holds a GPU Moving a running guest with an attached accelerator requires capturing device state, channel state, and page state in an order that leaves neither side with a stale view.