作者: Jason Riedy , Eric Hein , Tom Conte , Jeffrey Young , Srinivas Eswar
DOI: 10.1109/IPDPSW.2018.00097
关键词:
摘要: The Emu Chick is a prototype system designed around the concept of migratory memory-side processing. Rather than transferring large amounts data across power-hungry, high-latency interconnects, moves lightweight thread contexts to near-memory cores before beginning each memory read. current hardware uses FPGAs implement cache-less "Gossamer" for doing computational work and stationary core run basic operating functions migrate threads between nodes. In this initial characterization Chick, we study bandwidth characteristics through benchmarks like STREAM, pointer chasing, sparse matrix vector multiply. We compare architectural simulation Intel Xeon-based platforms. While it difficult accurately with existing systems, our evaluation demonstrates that available more efficiently traditional, cache-based architecture. Moreover, provides stable, predictable performance 80% utilization on random-access chasing benchmark weak locality.