Bulldozers, Termites, and Chainsaws: Finding a Gradual Path to Parallelism

Опубликовано 6 сентября 2016, 16:44

The transition to multicore is effectively a gamble that application programmers will start writing all of their code in parallel, or that workloads will change to make it simple to generate many threads. A key question for architects is the granularity of the processors on each multicore chip -- many small processors, a few medium-sized processors, or fewer large behemoths. In this talk, I will show that with an EDGE instruction set, this choice is a false choice. I will describe Composable Lightweight Processors, which allow larger logical processors to be synthesized out of smaller ones. This class of architectures treats large processors as a distributed system made up of many small pieces on an on-chip micronetwork. This approach creates the opportunity to have many intermediate design points between traditional uniprocessors and multiprocessors, blurring the distinction and permitting a more gradual path to fine-grain parallel execution than requiring full-blown parallel applications immediately.