Fast thread communication and synchronization mechanisms for a scalable single chip multiprocessor
Stephen W. Keckler, William J. Dally · DSpace@MIT (Massachusetts Institute of Technology) · 1998
Much of the improvement in computer performance over the last twenty y ears has come from faster transistors and architectural advances that increase parallelism.Smaller feature sizes have decreased the transistor switching time but at the same time increased the resistance of interconnect wires, resulting in slower signal transmission in on-chip wiring.Since future chips will have more silicon area and include more execution units, a much larger demand for parallelism is emerging.However, the increased signi cance of wire delay will require monolithic components, such a s processors and caches, to be small and that the communication wires connecting them be short.Computer systems typically exploit concurrency using either instruction level parallelism (ILP) or coarse-grain parallel threads running on a multiprocessor.This thesis proposes mechanisms for exploiting on-chip parallelism at a ne grain to bridge the gap between ILP and coarse-grain multiprocessing.Fast interprocessor communication and synchronization enables the use of tasks with run lengths as small as 10 cycles.At the same time, these interaction mechanisms are less susceptible than conventional microprocessor designs to longer wire delays imminent in future silicon process technologies.As ne-grain parallelism is orthogonal to ILP and coarse-grain threads, it complements both methods and provides an opportunity for greater speedup.This thesis presents the architecture and implementation of the MIT Multi-ALU Processor (MAP), a 5 million transistor custom VLSI microprocessor chip.The MAP architecture incorporates 9 function units, split into 3 independent processors.The processors communicate via interprocessor register writes and synchronize using a hardware barrier instruction.These integrated mechanisms allow threads to communicate 10 times faster and synchronize 60 times faster than using a shared on-chip cache.The fast interprocessor interaction enables the MAP to exploit both instruction-level parallelism and ne-grain thread level parallelism.On a suite of applications, speedups of 1.2{2.4 are achieved using ne-grain threads on a 3-processor MAP chip.