PIMP My Many-Core: Pipeline-Integrated Message Passing

To improve the scalability, several many-core architectures use message passing instead of shared memory accesses for communication. Unfortunately, Direct Memory Access (DMA) transfers in a shared address space are usually used to emulate message passing, which entails a lot of overhead and thwarts the advantages of message passing. Recently proposed register-level message passing alternatives use special instructions to send the contents of a single register to another core. The reduced communication overhead and architectural simplicity lead to good many-core scalability. After investigating several other approaches in terms of hardware complexity and throughput overhead, we recommend a small instruction set extension to enable register-level message passing at minimal hardware costs and describe its integration into a classical five stage RISC-V pipeline.