异构多核SoC处理器内部存储架构优化

doi:10.16180/j.cnki.issn1007-7820.2022.09.007

摘要/Abstract

摘要：

异构多核技术的发展使微处理器的性能有了较大提升,而处理器与外部存储器之间的带宽差异限制了处理器的性能发挥,“存储墙”问题日益严重。针对一种用于高密度计算的异构多核SoC系统,文中提出了一套存储设计方案。该方案通过复用一些长时间闲置的本地空闲存储资源作为二级共享缓存来增加访存带宽,减少访问外部存储频率。分布式高速共享二级缓存结合多路并行访问外部存储的层次化存储结构,缓解了系统处理数据与外部存储器间的速度差异,提高了数据的存取效率,优化了系统的性能。综合资源消耗和计算效率,文中所提设计相比普通二级缓存节约了69.36%的片上SRAM资源,相比无缓存结构提高了41.2%的加速比,整体任务计算时间平均减少了约40.6%。

关键词: 异构多核, 存储墙, 复用, 多路并行, 层次化存储, 二级缓存, 分布式, 外部存储器

Abstract:

The performance of microprocessors has been greatly improved by the development of heterogeneous multi-core technology. The bandwidth difference between the processor and external memory severely limits the performance of the processor, and the "Memory Wall" problem is becoming increasingly serious. For a heterogeneous multi-core SoC system in high-density computing, this study proposes a set of memory design scheme. The solution increases memory access bandwidth and reduces the frequency of accessing external memory by reusing some local free memory resources that have been idle for a long time as shared L2 cache. Meanwhile, the distributed high-speed shared L2 cache combined with the hierarchical storage structure of multi-channel parallel access to external storage alleviates the speed difference between system processing data and external storage, improves data access efficiency, and optimizes system performance. In terms of resource consumption and computing efficiency, the proposed design saves 69.36% of on-chip SRAM resources compared with ordinary L2 cache, provides 41.2% speedup ratio compared with non-cache structure, and reduces the overall task calculation time by about 40.6% on average.

Key words: heterogeneous multicore, memory wall, multiplexing, multiplexed parallelism, hierarchical memory, L2 cache, distributed, external memory

中图分类号:

TN47

张玄,张多利,宋宇鲲. 异构多核SoC处理器内部存储架构优化[J]. 电子科技, 2022, 35(9): 44-51.

ZHANG Xuan,ZHANG Duoli,SONG Yukun. Optimization of the Internal Memory Architecture of Heterogeneous Multi-Core SoC Processors[J]. Electronic Science and Technology, 2022, 35(9): 44-51.

图/表 15

图1

图2

表1

图3

图4

图5

图6

表2

表3

图7

图8

图9

图10

图11

图12

参考文献 18

[1]	Wang X L, Qu D M, Song Y K, et al. Design and implementation of homogeneous multi-core system[C]. Guiyang: Proceedings of IEEE the Twelfth International Conference on ASIC, 2017.
[2]	蒋林, 刘鹏, 山蕊, 等. 阵列处理器分布式存储的簇内全访问结构设计[J]. 西安科技大学学报, 2018, 38(4):656-662.
	Jiang Lin, Liu Peng, Shan Rui, et al. Design of intra-cluster full-switch architecture for distributed storage[J]. Journal of Xi'an University of Science and Technology, 2018, 38(4):656-662.
[3]	沈鹏程. 基于SystemC的多通道DDR控制器建模设计[D]. 南京: 南京大学, 2018.
	Sheng Pengcheng. Research on modeling of multi-channel DDR SDRAM controller based on SystemC[D]. Nanjing: Nanjing University, 2018.
[4]	Subha S. A reconfigurable cache architecture[C]. Bhubaneswar: Proceedings of the International Conference on High Performance Computing and Applications, 2014.
[5]	山蕊, 沈绪榜, 蒋林, 等. 面向阵列处理器的分布式共享存储结构设计[J]. 北京邮电大学学报, 2017, 40(4):9-15.
	Shan Rui, Shen Xubang, Jiang Lin, et al. Design of distributed shared memory structure for array processor[J]. Journal of Beijing University of Posts and Telecommunications, 2017, 40(4):9-15.
[6]	刘有耀, 张园, 山蕊. 阵列处理器分布式Cache的局部优先访问结构设计[J]. 计算机工程与科学, 2020, 42(4):580-587.
	Liu Youyao, Zhang Yuan, Shan Rui. An intra-cluster local-priority efficient-access switch in distributed Cache[J]. Computer Engineering & Science, 2020, 42(4):580-587.
[7]	张多利, 张宇, 宋宇鲲, 等. 一种多核SoC中基于Cache机制的存储结构设计[J]. 微电子学与计算机, 2017, 34(10):26-31.
	Zhang Duoli, Zhang Yu, Song Yukun, et al. A storage structure design based on the Cache mechanism in a multi-SoC[J]. Microelectronics & Computer, 2017, 34(10):26-31.
[8]	Wang Z Z, Wang C H, Yuan Z A, et al. A Reconfigurable Cache structure design for multi-core SOC[C]. Chengdu: Proceedings of the IEEE International Conference on Integrated Circuits,Technologies and Applications, 2019.
[9]	肖瑞瑾. 多核处理器层次化存储体系研究[D]. 上海: 复旦大学, 2012.
	Xiao Ruijin. Research on hierarchical storage system of multi-core processor[D]. Shanghai: Fudan University, 2012.
[10]	袁亚鹏. 面向异构多核系统的层次化存储结构设计与优化[D]. 合肥: 合肥工业大学, 2018.
	Yuan Yapeng. Design and optimization of hierarchical storage structure for heterogeneous multi-core system[D]. Hefei: Hefei University of Technology, 2018.
[11]	Zhang D, Shen X L, Song Y K. The implementation of large FFT convolution on heterogeneous multicore programmable system[C]. Chengdu: Proceedings of the International Conference on Integrated Circuits and Microsystems, 2016.
[12]	Cheng X H, Tan C P, Zhang Y. Research on optimization of DAG task scheduling model based on heterogeneous multicore processor[C]. Chongqing: Proceedings of IEEE the Third Advanced Information Technology, Electronic and Automation Control Conference, 2018.
[13]	迟凯. Ad Hoc网络中基于信息论的可信模型研究[J]. 电子科技, 2018, 31(9):21-24.
	Chi Kai. Research of trust evaluation model based on information theoretic framework in Ad Hoc network[J]. Electronic Science and Technology, 2018, 31(9): 21-24.
[14]	孙远昕, 秦水介. LTE上行DFT硬件加速器的设计[J]. 电子科技, 2018, 31(4):52-54.
	Sun Yuanxin, Qin Shuijie. Design of DFT hardware accelerator used in LTE uplink[J]. Electronic Science and Technology, 2018, 31(4):52-54.
[15]	Park C H, Heo T, Jeong J, et al. Hybrid TLB coalescing: Improving TLB translation coverage under diverse fragmented memory allocations[C]. Toronto: Proceedings of the Forty-fourth Annual International Symposium on Computer Architecture, 2017.
[16]	Awad A, Basu A, Blagodurov S, et al. Avoiding TLB shootdowns through self-invalidating TLB entries[C]. Portland: Proceedings of the Twenty-sixth International Conference on Parallel Architectures and Compilation Techniques, 2017.
[17]	Kumar R. Analysis of shape alignment using Euclidean and Manhattan distance metrics[C]. Bhopal: Proceedings of the International Conference on Recent Innovations in Signal processing and Embedded Systems, 2017.
[18]	Sasaki S, Yasuda M, Mattausch H J. Digital associative memory for word-parrallel manhattan-distance-based vector quantization[C]. Bordeaux: Proceedings of the ESSCIRC, 2012.

任务类型	占用数据通道资源	功能类型
Data_type0_mc	网络与外部主存侧数据通道	主控指令信息
Data_type0_rcu	网络与外部主存侧数据通道	RCU指令执信息
Data_type0	网络与外部主存侧数据通道	普通数据读出
Data_type1	网络与外部主存侧数据通道	普通类型数据广播
Data_type2	网络与外部主存侧数据通道	普通类型数据写入
Data_type3	网络与二级缓存侧数据通道	读缓存数据
Data_type4	网络与二级缓存侧数据通道	写缓存数据

实验组号	实验名称	RCA	实验特征
实验1	矩阵转置	0	没有计算需求,DDR无空闲带宽
实验2	简单运算	<1	FFT为共享存储,DDR有空闲带宽
实验3	复杂运算	>1FFT	为私有存储,DDR有空闲带宽

逻辑资源	本文	文献[7]	文献[10]
Slice LUTs	85 525(6.99%)	70 146(5.74%)	96 555(7.90%)
Slice register	112 563(4.60%)	104 133(4.26%)	101 316(4.10%)
Block RAM36	110(8.50%)	359(27.79%)	110(8.50%)