An FPGA Implementation of a Distributed Virtual Machine

Abstract

An expression in a functional programming language can be compiled into a massively redundant, spatially distributed, concurrent computation called a distributed virtual machine (DVM). A DVM is comprised of bytecodes reified as actors undergoing diffusion on a two-dimensional grid communicating via messages containing encapsulated virtual machine states (continuations). Because the semantics of expression evaluation are purely functional, DVMs can employ massive redundancy in the representation of the heap to help ensure that computations complete even when large areas of the physical host substrate have failed. Because they can be implemented as asynchronous circuits, DVMs also address the well known problem affecting traditional machine architectures implemented as integrated circuits, namely, clock networks consuming increasingly large fractions of area as device size increases. This paper describes the first hardware implementation of a DVM. This was accomplished by compiling a VHDL specification of a special purpose distributed memory multicomputer with a mesh interconnection network into a globally asynchronous, locally synchronous (GALS) circuit in an FPGA. Each independently clocked node combines a processor based on a virtual machine for compiled Scheme language programs, with just enough local memory to hold a single heap allocated object and a continuation.