Parallel Simulation of Many-core Processors: Integration of Research and Education
Tali Moreshet, Uzi Vishkin, Fuat Keceli · 2020
Abstract Parallel Simulation of Many-core Processors: Integration of Research and EducationAbstractProviding undergraduate students with an opportunity to experience academic research has po-tential impact on their future career choice. More specifically, we believe that having a positiveresearch experience can motivate women and minority undergraduate engineering students to pur-sue an engineering career path, academic or otherwise. In this paper, we describe an on-going,multiple-year research project, led by undergraduate female students, which incorporates researchand education in computer science and engineering (CS&E). Our research project involves many-cores, which are becoming increasingly popular in general-purpose computing. While most researchers agree that this requires introduction of parallelism tomainstream CS&E practice, and hence education, parallel programming difficulties remain an ob-stacle that it yet to be overcome. The eXplicit Multi-Threading (XMT) framework provides ageneral-purpose many-core architecture for fine-grained parallel programs that scales to hundreds,or even thousands of lightweight cores. XMT aims at improving single task execution time throughparallelism, and has been supported by significant evidence on ease-of-programming and compet-itive performance. The XMT platform consists of a proof-of-concept 64-core FPGA and ASIC prototypes and ahighly configurable cycle-accurate simulator (XMTSim), capable of modeling a target 1024-coreXMT. Our work aims to parallelize XMTSim, with several objectives: (i) Long-term theoreticalobjective: Establish that XMT is an effective self simulating machine; namely simulating effi-ciently the XMTSim code by XMTSim itself. Recall that one of the elegant features of Turingmachines was their ability to provide self-simulations, and this was taken as evidence for theirgeneral-purposeness. (ii) Mid-term applicable objective: Speed up a many-core simulator by par-allelizing its most time-consuming components, beginning with the main simulation bottleneck,the Interconnection Network (ICN). (iii) Short-term training objective: Implementing a parallelapplication suitable for the XMT environment in the form of a parallel ICN. Our preliminary results for the short-term objective already show that parallelizing the ICNobtains simulation speedups of x54 compared to the best serial implementation on a 64-core XMT.We also explore alternative parallel implementations and demonstrate their potential advantage. While undergraduates realize that their work cannot achieve a high-impact research goal alone,we find that they appreciate being able to contribute towards a high-impact goal. Namely, toestablish, for the first time, that a parallel architecture (XMT, in this case) is general-purpose. Asexplained above, we plan to to do so by showing that XMT meets a classic understanding of whatgeneral-purpose computing is.