Elastic threads on composable processors

Modern CMPs are designed to exploit both instruction-level parallelism within processors and threadlevel parallelism across processors, but the balance between the granularity of each processor and the number of processors must be chosen at design time. In this paper, we propose a microarchitecture that allows this balance to be dynamically adjusted. The microarchitecture, which implements the TRIPS instruction set, consists of a large number of fine-grained, single-issue processor cores. By changing a set of OS-visible configuration registers, the system software can aggregate multiple cores to form larger, more powerful processors, depending on the needs of the available threads. For instance, a 64-core chip could be configured as 64 1-wide processors, 1 64-wide processor, or any combination in between. We quantify the area and performance overheads associated with providing the capability to compose larger processors out of multiple small ones, find the distinct ideal configuration for each of several applications, and show the additional benefits gained by explicit compiler support for specific configurations. (a) 32 1-wide TFlex config. (b) 8-processor TFlex config. (c) One 32-wide TFlex config. P P P