When Application-Specific ISA Meets FPGAs: A Multi-LayerVirtualization Framework for Heterogeneous Cloud FPGAs

Abstract

The paper is the first to propose a new virtualization mechanism that can natively support scale-out acceleration across heterogeneous cloud FPGAs and demonstrate its effectiveness on a custom FPGA cluster with heterogeneous FPGA resources. Cloud FPGA virtualization mainly targets homogeneous FPGAs in the previous works. The key contribution of this work is to introduce a new system abstraction that can serve as an indirection layer to bridge high-level Application-specific (AS) ISA and low-level Hardware-specific (HS) abstraction. A set of custom tools is developed to implement the proposed virtualization mechanism with high mapping quality. Additional optimization technique and system support are also provided to maximally hide the inter-FPGA communication latency when exploiting scale-out acceleration using multiple heterogeneous FPGAs.