G-View: The Software Performance Monitor For Godson-T Many-Core Architecture: Design And A Study Of Test Algorithms

Godson-T is the many-core system architecture suitable for ultra-deep submicron technology under development by Institute of Computing Technology,Chinese Academy of Sciences.Nowadays,on-chip performance counters have been widely used for performance analysis of applications and for power analysis of chips,but only if supported by applicable software,could hardware performance counters work appropriately.In order to support the need for Godson-T users to analyze and optimize applications,and that for Godson-T developers to optimize Godson-T performance and power,we design and implement G-View,the software performance monitor based on Godson-T simulator and FPGA implementation.On the other hand,parallel models and parallel algorithms,acting as a criterion for parallel system architecture,have always been hot spots.In order to provide test cases for integration testing of G-View and Godson-T,we also propose Godson-T parallel model and Broadcast algorithm,and analyze and implement several non-numeric parallel algorithms based on the algorithm.We do the following work:1,we do requirement analysis for the real-time performance monitor based on Godson-T architecture.2,we design and implement the real-time performance monitor G-View of two versions:one for Godson-T simulator and the other for Godson-T FPGA.Round Robin algorithm is used to obtain values from Godson-T performance counters; loosely-coupled interactions between G-View and Godson-T and between G-View modules are achieved;GUI and Performance Analyzer modules are merged in one thread.3,we analyze and implement Parallel Select,Parallel Merge and Parallel Sort based on Godson-T memory consistency model,providing test cases for system testing of G-View and Godson-T.4,we propose Extended Parallel External Memory(EPEM)model,and the basic parallel algorithm Broadcast.We also prove optimization of the algorithm theoretically and experimentally.The Round Robin strategy can efficiently monitor Godson-T performance;the loosely-coupled design between G-View and interaction object supports changes of the object;loosely-coupled design between G-View models supports Godson-T updating; Extended PEM model based Broadcast algorithm optimizes performance of test cases, and solves the problem that there is no parallel model for the state-of-art Exclusive Read Exclusive Write(EREW)shared memory system architectures;LU decomposition invoking this algorithm can reach a performance gain of 42.9%under 64 cores.