Artifacts of AvA (Accelerated Virtualization of Accelerators) in ASPLOS'20

This is the artifacts of the paper titled "AvA: Accelerated Virtualization of Accelerators" which will appear in ASPLOS'20. Abstract: Applications are migrating en masse to the cloud, while accelerators such as GPUs, TPUs, and FPGAs proliferate in the wake of Moore's Law. These trends are in conflict: cloud applications run on virtual platforms, but existing virtualization techniques have not provided production-ready solutions for accelerators. As a result, cloud providers expose accelerators by dedicating physical devices to individual guests. Multi-tenancy and consolidation are lost as a consequence. We propose automatic generation of virtual accelerator stacks to address the fundamental limitations of existing virtualization techniques. AvA provides automated construction of support for hypervisor-mediated accelerator sharing among mutually distrustful VMs. AvA combines a DSL for describing accelerator APIs and sharing policies, a device-agnostic runtime, and tools to generate and deploy accelerator-specific stack components such as guest libraries and API servers. AvA uses a novel technique called Hypervisor Interposed Remote Acceleration (HIRA) that retains hypervisor interposition for efficient policy enforcement. We used AvA to virtualize ten accelerators and framework APIs, including six for which no virtualization support has been previously explored. Our evaluation shows that AvA can provide near-native performance and enforce resource sharing policies that are not possible with current techniques such as SR-IOV and user-level API remoting, all with orders of magnitude lower programming effort than required to construct hand-built virtualization support.The project is actively maintained in https://github.com/utcs-scea/ava