Junco: Live virtual machine migration across CPU and FPGA hardware 20-000000800
Abstract
The project proposes Junco – an infrastructure hardware/software stack that enables migrating virtual machines from server-class machines, at run-time, to FPGAs that are dynamically programmed to support their execution. The project envisions this capability to significantly enhance fault-resilience of mission-critical systems by offering another knob in the faultmanagement design space to cope with failures. While live virtual machine (VM) migration has been extensively studied and used in cloud infrastructures, among other settings, for failure resilience, they are limited to CPU-based machines, often to those with the same instruction sets or micro-architectural properties. This project extends VM migration seamlessly across CPUs and FPGAs, which requires overcoming significant hardware and software challenges, including reconfiguring FPGAs to house VMs with no loss of functionality, acceptable performance, and continued isolation. The project proposes to overcome these challenges through innovations across the stack – FPGAs, hypervisors, compilers, and run-times – that provide hypervisor-like interfaces and unified virtual memory for FPGAs to support VM execution, application-specific MMU and TLB implementations on FPGAs, decomposition of VMs into unikernel VMs that are appropriate for CPU and FPGA executions, and fault- and resource management policies that orchestrate their seamless execution. The project will contribute to Ship system survivability by offering another mechanism for managing component failures through live migration of VMs across CPU-FPGAinfrastructures.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 20, 2021
- Source ID
- N000142112523
Entities
People
- Binoy Ravindran
Organizations
- Office of Naval Research
- United States Navy
- Virginia Tech