Research On Heterogeneous Multicore Platform Oriented Efficient Speculative Parallelizing Technology

Compared with traditional single-core processors,multi-core processors are with lower power consumption and higher efficiency.After several years' continuous development,multi-core processors have become more and more mature,and have been widely used in the field of high performance computing,desktop computing and mobile computing etc.Meanwhile,multi-core processors are also turning to heterogeneous.However,heterogeneous multi-core architecture in a variety of processor convergence technology is becoming increasingly complex and diversified,and hence brings difficulty for the study of heterogeneous multi-core platform technology.First of all,unlike the architecture that has rapidly developed,the research of programming model for heterogeneous multi-core platform is still at its early stage,which means that we still can not take full advantage of computing resource of heterogeneous multi-core platform.In addition,as the number of cores on chip increases and frequency of processor core grows,the power consumption of heterogeneous multi-core processor has become increasingly prominent.To tackle these issues,this article aims at exploring the programming model and power consumption optimization for heterogeneous multi-core platform,and the main contributions of this work are as follows:1)Three different comparative heterogeneous multi-core platforms are selected for comprehensive analysis.After studying the performance of Intel MIC,Micron AP and AMD APU regarding speculative parallel technique,we have the following observations.(1)Since the Intel MIC is based on the X86 architecture,the traditional parallelism mechanism for the isomorphic system can be well migrated to the MIC.(2)Although the AP can be used to accelerate the specific computing tasks to obtain a high speedup,the AP product is still not mature and some interfaces are still not available.Therefore,currently the AP still cannot support programmable heterogeneous computing.(3)The CPU + GPU heterogeneous system.e.g.,AMD APU,owns huge resources to support thread-level parallelism.However,how to design GPU-oriented speculative parallel mechanism remains a meaningful research direction.Therefore,after the comprehensive analysis,this paper holds that it is of great research value to explore the speculative parallel technology for CPU + GPU heterogeneous multi-core platform.2)A GPU-oriented software-based thread-level speculative parallel model called STLS is introduced.The main idea is to adopt the code transformation to force the serial execution of data that has dependence;Secondly,using the speculation to optimize the control divergence,so that threads can be executed in parallel as much as possible.In order to implement the STLS model and improve its performance,two kinds of basic data are summarized from the application program,and a simple dynamic data correlation detection method is designed according to their characteristics.This work also puts forward the speculative data management mechanism of STLS system,and analyzes its performance overhead.Moreover,to reduce the performance loss caused by thread speculation error,the Warp data in Warp is reduced by moving the Warp memory in one or more of several threads related to data.The test results for typical applications show that the STLS system can provide effective performance improvement.3)A power consumption prediction model is proposed for integrated GPU.The 15 regression factors in the regression model were identified by classification analysis and the Rodinia Benchmark Suite was chosen as the benchmark of the test.In addition,in order to be able to accurately measure the model data,we propose a Kernel extension mechanism.By extending the number of executions of all kernels,the run time of each kernel is increased to meet the requirements of the CodeXL sampling period.We then used the SPSS method to evaluate the regression model by extracting the power consumption data generated by the underlying running program.The test results show that the model has good linear correlation and the overall accuracy of the model is high.In order to reduce the delay of power consumption prediction model,we further explores the method of exploring simplified power consumption model.4)The dynamic power regulation technique of STLS system under power-limited condition is studied.We first analyze two power consumption adjustment methods,i.e.,DVFS and working mode switching,for the STLS system.The former requires the GPU to support DVFS,which requires the GPU to support thread-level Power Capping.In this paper,the Power Capping strategy STLS-DPA is designed for the STLS system.In each adjustment cycle,the strategy can select the most suitable VF level and STLS operation mode according to the STLS system power consumption value predicted by the GPU power consumption prediction model,and adjust the STF system VF level and operating mode.The experiment results show that compared with the test strategy,STLS-DPA strategy for STLS system occurs power consumption similar to the target power consumption,and moreover has faster adjustment speed.