Due to advances in semiconductor techniques, many-core processors have been widely used in high performance computing. However, many applications still cannot be carried out efficiently due to the memory wall, which h...Due to advances in semiconductor techniques, many-core processors have been widely used in high performance computing. However, many applications still cannot be carried out efficiently due to the memory wall, which has become a bottleneck in many-core processors. In this paper, we present a novel heterogeneous many-core processor architecture named deeply fused many-core （DFMC） for high performance computing systems. DFMC integrates management processing ele- ments （MPEs） and computing processing elements （CPEs）, which are heterogeneous processor cores for different application features with a unified ISA （instruction set architecture）, a unified execution model, and share-memory that supports cache coherence. The DFMC processor can alleviate the memory wall problem by combining a series of cooperative computing techniques of CPEs, such as multi-pattern data stream transfer, efficient register-level communication mechanism, and fast hardware synchronization technique. These techniques are able to improve on-chip data reuse and optimize memory access performance. This paper illustrates an implementation of a full system prototype based on FPGA with four MPEs and 256 CPEs. Our experimental results show that the effect of the cooperative computing techniques of CPEs is significant, with DGEMM （double-precision matrix multiplication） achieving an efficiency of 94%, FFT （fast Fourier transform） obtaining a performance of 207 GFLOPS and FDTD （finite-difference time-domain） obtaining a performance of 27 GFLOPS.展开更多
As semiconductor technology advances, there will be billions of transistors on a single chip. Chip many-core processors are emerging to take advantage of these greater transistor densities to deliver greater performan...As semiconductor technology advances, there will be billions of transistors on a single chip. Chip many-core processors are emerging to take advantage of these greater transistor densities to deliver greater performance. Effective fault tolerance techniques are essential to improve the yield of such complex chips. In this paper, a core-level redundancy scheme called N+M is proposed to improve N-core processors’ yield by providing M spare cores. In such architecture, topology is an important factor because it greatly affects the processors’ performance. The concept of logical topology and a topology reconfiguration problem are introduced, which is able to transparently provide target topology with lowest performance degradation as the presence of faulty cores on-chip. A row rippling and column stealing (RRCS) algorithm is also proposed. Results show that PRCS can give solutions with average 13.8% degradation with negligible computing time.展开更多
Purpose–The purpose of this paper is to propose a fault-tolerant technology for increasing the durability of application programs when evolutionary computation is performed by fast parallel processing on many-core pr...Purpose–The purpose of this paper is to propose a fault-tolerant technology for increasing the durability of application programs when evolutionary computation is performed by fast parallel processing on many-core processors such as graphics processing units(GPUs)and multi-core processors(MCPs).Design/methodology/approach–For distributed genetic algorithm(GA)models,the paper proposes a method where an island’s ID number is added to the header of data transferred by this island for use in fault detection.Findings–The paper has shown that the processing time of the proposed idea is practically negligible in applications and also shown that an optimal solution can be obtained even with a single stuck-at fault or a transient fault,and that increasing the number of parallel threads makes the system less susceptible to faults.Originality/value–The study described in this paper is a new approach to increase the sustainability of application program using distributed GA on GPUs and MCPs.展开更多
With the advent of the big data era,the amounts of sampling data and the dimensions of data features are rapidly growing.It is highly desired to enable fast and efficient clustering of unlabeled samples based on featu...With the advent of the big data era,the amounts of sampling data and the dimensions of data features are rapidly growing.It is highly desired to enable fast and efficient clustering of unlabeled samples based on feature similarities. As a fundamental primitive for data clustering,the k-means operation is receiving increasingly more attentions today.To achieve high performance k-means computations on modern multi-core/many-core systems,we propose a matrix-based fused framework that can achieve high performance by conducting computations on a distance matrix and at the same time can improve the memory reuse through the fusion of the distance-matrix computation and the nearest centroids reduction.We implement and optimize the parallel k-means algorithm on the SW26010 many-core processor,which is the major horsepower of Sunway TaihuLight.In particular,we design a task mapping strategy for load-balanced task distribution,a data sharing scheme to reduce the memory footprint and a register blocking strategy to increase the data locality.Optimization techniques such as instruction reordering and double buffering are further applied to improve the sustained performance.Discussions on block-size tuning and performance modeling are also presented.We show by experiments on both randomly generated and real-world datasets that our parallel implementation of k-means on SW26010 can sustain a double-precision performance of over 348.1 Gflops,which is 46.9% of the peak performance and 84%of the theoretical performance upper bound on a single core group,and can achieve a nearly ideal scalability to the whole SW26010 processor of four core groups.Performance comparisons with the previous state-of-the-art on both CPU and GPU are also provided to show the superiority of our optimized k-means kernel.展开更多
The flourish of deep learning frameworks and hardware platforms has been demanding an efficient compiler that can shield the diversity in both software and hardware in order to provide application portability.Among th...The flourish of deep learning frameworks and hardware platforms has been demanding an efficient compiler that can shield the diversity in both software and hardware in order to provide application portability.Among the existing deep learning compilers,TVM is well known for its efficiency in code generation and optimization across diverse hardware devices.In the meanwhile,the Sunway many-core processor renders itself as a competitive candidate for its attractive computational power in both scientific computing and deep learning workloads.This paper combines the trends in these two directions.Specifically,we propose swTVM that extends the original TVM to support ahead-of-time compilation for architecture requiring cross-compilation such as Sunway.In addition,we leverage the architecture features during the compilation such as core group for massive parallelism,DMA for high bandwidth memory transfer and local device memory for data locality,in order to generate efficient codes for deep learning workloads on Sunway.The experiment results show that the codes generated by swTVM achieve 1.79x improvement of inference latency on average compared to the state-of-the-art deep learning framework on Sunway,across eight representative benchmarks.This work is the first attempt from the compiler perspective to bridge the gap of deep learning and Sunway processor particularly with productivity and efficiency in mind.We believe this work will encourage more people to embrace the power of deep learning and Sunwaymany-coreprocessor.展开更多
Recent architectures of multi-core systems may have a relatively large number of cores that typically ranges from tens to hundreds;therefore called many-core systems.Such systems require an efficient interconnection n...Recent architectures of multi-core systems may have a relatively large number of cores that typically ranges from tens to hundreds;therefore called many-core systems.Such systems require an efficient interconnection network that tries to address two major problems.First,the overhead of power and area cost and its effect on scalability.Second,high access latency is caused by multiple cores’simultaneous accesses of the same shared module.This paper presents an interconnection scheme called N-conjugate Shuffle Clusters(NCSC)based on multi-core multicluster architecture to reduce the overhead of the just mentioned problems.NCSC eliminated the need for router devices and their complexity and hence reduced the power and area costs.It also resigned and distributed the shared caches across the interconnection network to increase the ability for simultaneous access and hence reduce the access latency.For intra-cluster communication,Multi-port Content Addressable Memory(MPCAM)is used.The experimental results using four clusters and four cores each indicated that the average access latency for a write process is 1.14785±0.04532 ns which is nearly equal to the latency of a write operation in MPCAM.Moreover,it was demonstrated that the average read latency within a cluster is 1.26226±0.090591 ns and around 1.92738±0.139588 ns for read access between cores from different clusters.展开更多
In this paper,a typical experiment is carried out based on a high-resolution air-sea coupled model,namely,the coupled ocean-atmosphere-wave-sediment transport(COAWST)model,on both heterogeneous many-core(SW)and homoge...In this paper,a typical experiment is carried out based on a high-resolution air-sea coupled model,namely,the coupled ocean-atmosphere-wave-sediment transport(COAWST)model,on both heterogeneous many-core(SW)and homogenous multicore(Intel)supercomputing platforms.We construct a hindcast of Typhoon Lekima on both the SW and Intel platforms,compare the simulation results between these two platforms and compare the key elements of the atmospheric and ocean modules to reanalysis data.The comparative experiment in this typhoon case indicates that the domestic many-core computing platform and general cluster yield almost no differences in the simulated typhoon path and intensity,and the differences in surface pressure(PSFC)in the WRF model and sea surface temperature(SST)in the short-range forecast are very small,whereas a major difference can be identified at high latitudes after the first 10 days.Further heat budget analysis verifies that the differences in SST after 10 days are mainly caused by shortwave radiation variations,as influenced by subsequently generated typhoons in the system.These typhoons generated in the hindcast after the first 10 days attain obviously different trajectories between the two platforms.展开更多
With the development of computer technology, network bandwidth and network traffic continue to increase. Considering the large data flow, it is imperative to perform inspection effectively on network packets. In order...With the development of computer technology, network bandwidth and network traffic continue to increase. Considering the large data flow, it is imperative to perform inspection effectively on network packets. In order to find a solution of deep packet inspection which can appropriate to the current network environment, this paper built a deep packet inspection system based on many-core platform, and in this way, verified the feasibility to implement a deep packet inspection system under many-core platform with both high performance and low consumption. After testing and analysis of the system performance, it has been found that the deep packet inspection based on many-core platform TILE_Gx36 [1] [2] can process network traffic of which the bandwidth reaches up to 4 Gbps. To a certain extent, the performance has improved compared to most deep packet inspection system based on X86 platform at present.展开更多
Simulators are generally used during the design of computer architectures. Typically, different simulators with different levels of complexity, speed and accuracy are used. However, for early design space exploration,...Simulators are generally used during the design of computer architectures. Typically, different simulators with different levels of complexity, speed and accuracy are used. However, for early design space exploration, simulators with less complexity, high simulation speed and reasonable accuracy are desired. It is also required that these simulators have a short development time and that changes in the design require less effort in the implementation in order to perform experiments and see the effects of changes in the design. These simulators are termed high-level simulators in the context of computer architecture. In this paper, we present multiple levels of abstractions in a high-level simulation of a general-purpose many-core system, where the objective of every level is to improve the accuracy in simulation without significantly affecting the complexity and simulation speed.展开更多
A team of researchers from the University of Science and Technology of China(USTC)of the Chinese Academy of Sciences(CAS)and its partners have made significant advancements in random quantum circuit sampling with Zuch...A team of researchers from the University of Science and Technology of China(USTC)of the Chinese Academy of Sciences(CAS)and its partners have made significant advancements in random quantum circuit sampling with Zuchongzhi-3,a superconducting quantum computing prototype featuring 105 qubits and 182 couplers.展开更多
目前,多核实时系统中同步任务的节能调度研究主要针对的是同构多核处理器平台,而异构多核处理器架构能够更有效地发挥系统性能。将现有的研究直接应用于异构多核系统,在保证可调度性的情况下会导致能耗变高。对此,通过使用动态电压与频...目前,多核实时系统中同步任务的节能调度研究主要针对的是同构多核处理器平台,而异构多核处理器架构能够更有效地发挥系统性能。将现有的研究直接应用于异构多核系统,在保证可调度性的情况下会导致能耗变高。对此,通过使用动态电压与频率调节(Dynamic Voltage Frequency Scaling,DVFS)技术,研究异构多核实时系统中基于任务同步的节能调度问题,提出同步感知的最大能耗节省优先算法(Synchronization Aware-Largest Energy Saved First,SA-LESF)。该算法针对所有任务的速度配置进行迭代优化,直至所有任务均达到其最大限度节能的速度配置。此外,进一步提出基于动态松弛时间回收的同步感知最大能耗节省优先算法(Synchronization Aware-Largest Energy Saved First with Dynamic Reclamation,SA-LESF-DR)。该算法在保证实时任务可调度的同时,实施相应的回收策略,进一步降低系统能耗。实验结果表明,SA-LESF与SA-LESF-DR算法在能耗表现上具有优势,在相同任务集下,相比其他算法可节省高达30%的能耗。展开更多
This paper presents a smart compensation system based on MCA7707 (a kind of signal processor). The li near errors and high order errors of a sensor (especially piezoresistive sensor) can be corrected by using this s...This paper presents a smart compensation system based on MCA7707 (a kind of signal processor). The li near errors and high order errors of a sensor (especially piezoresistive sensor) can be corrected by using this system. It can optimize the process of piezoresi stive sensor calibration and compensation, then, a total error factor within 0.2 % of the sensor′s repeatability errors is obtained. Data are recorded and coeff icients are determined automatically by this system, thus, the sensor compensati on is simplified greatly. For operating easily, a wizard compensation program is designed to correct every error and to get the optimum compensation.展开更多
文摘Due to advances in semiconductor techniques, many-core processors have been widely used in high performance computing. However, many applications still cannot be carried out efficiently due to the memory wall, which has become a bottleneck in many-core processors. In this paper, we present a novel heterogeneous many-core processor architecture named deeply fused many-core （DFMC） for high performance computing systems. DFMC integrates management processing ele- ments （MPEs） and computing processing elements （CPEs）, which are heterogeneous processor cores for different application features with a unified ISA （instruction set architecture）, a unified execution model, and share-memory that supports cache coherence. The DFMC processor can alleviate the memory wall problem by combining a series of cooperative computing techniques of CPEs, such as multi-pattern data stream transfer, efficient register-level communication mechanism, and fast hardware synchronization technique. These techniques are able to improve on-chip data reuse and optimize memory access performance. This paper illustrates an implementation of a full system prototype based on FPGA with four MPEs and 256 CPEs. Our experimental results show that the effect of the cooperative computing techniques of CPEs is significant, with DGEMM （double-precision matrix multiplication） achieving an efficiency of 94%, FFT （fast Fourier transform） obtaining a performance of 207 GFLOPS and FDTD （finite-difference time-domain） obtaining a performance of 27 GFLOPS.
基金the National Natural Science Foundation of China (Nos. 60633060, 60606008, and 60576031)the National Key Basic Research and Development (973) Program of China (973)(Nos. 2005CB321604 and 2005CB321605)the fund of Chinese Academy of Sciences (No. 20074010) due to the President Scholarship
文摘As semiconductor technology advances, there will be billions of transistors on a single chip. Chip many-core processors are emerging to take advantage of these greater transistor densities to deliver greater performance. Effective fault tolerance techniques are essential to improve the yield of such complex chips. In this paper, a core-level redundancy scheme called N+M is proposed to improve N-core processors’ yield by providing M spare cores. In such architecture, topology is an important factor because it greatly affects the processors’ performance. The concept of logical topology and a topology reconfiguration problem are introduced, which is able to transparently provide target topology with lowest performance degradation as the presence of faulty cores on-chip. A row rippling and column stealing (RRCS) algorithm is also proposed. Results show that PRCS can give solutions with average 13.8% degradation with negligible computing time.
文摘Purpose–The purpose of this paper is to propose a fault-tolerant technology for increasing the durability of application programs when evolutionary computation is performed by fast parallel processing on many-core processors such as graphics processing units(GPUs)and multi-core processors(MCPs).Design/methodology/approach–For distributed genetic algorithm(GA)models,the paper proposes a method where an island’s ID number is added to the header of data transferred by this island for use in fault detection.Findings–The paper has shown that the processing time of the proposed idea is practically negligible in applications and also shown that an optimal solution can be obtained even with a single stuck-at fault or a transient fault,and that increasing the number of parallel threads makes the system less susceptible to faults.Originality/value–The study described in this paper is a new approach to increase the sustainability of application program using distributed GA on GPUs and MCPs.
基金the National Key Research and Development Plan of China under Grant No.2016YFB0200603the National Natural Science Foundation of China under Grant No.91530323the Beijing Natural Science Foundation of China under Grant No.JQ18001.
文摘With the advent of the big data era,the amounts of sampling data and the dimensions of data features are rapidly growing.It is highly desired to enable fast and efficient clustering of unlabeled samples based on feature similarities. As a fundamental primitive for data clustering,the k-means operation is receiving increasingly more attentions today.To achieve high performance k-means computations on modern multi-core/many-core systems,we propose a matrix-based fused framework that can achieve high performance by conducting computations on a distance matrix and at the same time can improve the memory reuse through the fusion of the distance-matrix computation and the nearest centroids reduction.We implement and optimize the parallel k-means algorithm on the SW26010 many-core processor,which is the major horsepower of Sunway TaihuLight.In particular,we design a task mapping strategy for load-balanced task distribution,a data sharing scheme to reduce the memory footprint and a register blocking strategy to increase the data locality.Optimization techniques such as instruction reordering and double buffering are further applied to improve the sustained performance.Discussions on block-size tuning and performance modeling are also presented.We show by experiments on both randomly generated and real-world datasets that our parallel implementation of k-means on SW26010 can sustain a double-precision performance of over 348.1 Gflops,which is 46.9% of the peak performance and 84%of the theoretical performance upper bound on a single core group,and can achieve a nearly ideal scalability to the whole SW26010 processor of four core groups.Performance comparisons with the previous state-of-the-art on both CPU and GPU are also provided to show the superiority of our optimized k-means kernel.
基金supported by the National Key Research and Development Program of China (No.2020YFB1506703)the National Natural Science Foundation of China (Grant Nos.62072018 and 61732002)+1 种基金the State Key Laboratory of Software Development Environment (No.SKLSDE-2021ZX-06)the Fundamental Research Funds for the Central Universities。
文摘The flourish of deep learning frameworks and hardware platforms has been demanding an efficient compiler that can shield the diversity in both software and hardware in order to provide application portability.Among the existing deep learning compilers,TVM is well known for its efficiency in code generation and optimization across diverse hardware devices.In the meanwhile,the Sunway many-core processor renders itself as a competitive candidate for its attractive computational power in both scientific computing and deep learning workloads.This paper combines the trends in these two directions.Specifically,we propose swTVM that extends the original TVM to support ahead-of-time compilation for architecture requiring cross-compilation such as Sunway.In addition,we leverage the architecture features during the compilation such as core group for massive parallelism,DMA for high bandwidth memory transfer and local device memory for data locality,in order to generate efficient codes for deep learning workloads on Sunway.The experiment results show that the codes generated by swTVM achieve 1.79x improvement of inference latency on average compared to the state-of-the-art deep learning framework on Sunway,across eight representative benchmarks.This work is the first attempt from the compiler perspective to bridge the gap of deep learning and Sunway processor particularly with productivity and efficiency in mind.We believe this work will encourage more people to embrace the power of deep learning and Sunwaymany-coreprocessor.
文摘Recent architectures of multi-core systems may have a relatively large number of cores that typically ranges from tens to hundreds;therefore called many-core systems.Such systems require an efficient interconnection network that tries to address two major problems.First,the overhead of power and area cost and its effect on scalability.Second,high access latency is caused by multiple cores’simultaneous accesses of the same shared module.This paper presents an interconnection scheme called N-conjugate Shuffle Clusters(NCSC)based on multi-core multicluster architecture to reduce the overhead of the just mentioned problems.NCSC eliminated the need for router devices and their complexity and hence reduced the power and area costs.It also resigned and distributed the shared caches across the interconnection network to increase the ability for simultaneous access and hence reduce the access latency.For intra-cluster communication,Multi-port Content Addressable Memory(MPCAM)is used.The experimental results using four clusters and four cores each indicated that the average access latency for a write process is 1.14785±0.04532 ns which is nearly equal to the latency of a write operation in MPCAM.Moreover,it was demonstrated that the average read latency within a cluster is 1.26226±0.090591 ns and around 1.92738±0.139588 ns for read access between cores from different clusters.
基金This work is supported by the National Key Research and Development Plan program of the Ministry of Science and Technology of China(No.2016YFB0201100)Additionally,this work is supported by the National Laboratory for Marine Science and Technology(Qingdao)Major Project of the Aoshan Science and Technology Innovation Program(No.2018ASKJ01-04)the Open Fundation of Key Laboratory of Marine Science and Numerical Simulation,Ministry of Natural Resources(No.2021-YB-02).
文摘In this paper,a typical experiment is carried out based on a high-resolution air-sea coupled model,namely,the coupled ocean-atmosphere-wave-sediment transport(COAWST)model,on both heterogeneous many-core(SW)and homogenous multicore(Intel)supercomputing platforms.We construct a hindcast of Typhoon Lekima on both the SW and Intel platforms,compare the simulation results between these two platforms and compare the key elements of the atmospheric and ocean modules to reanalysis data.The comparative experiment in this typhoon case indicates that the domestic many-core computing platform and general cluster yield almost no differences in the simulated typhoon path and intensity,and the differences in surface pressure(PSFC)in the WRF model and sea surface temperature(SST)in the short-range forecast are very small,whereas a major difference can be identified at high latitudes after the first 10 days.Further heat budget analysis verifies that the differences in SST after 10 days are mainly caused by shortwave radiation variations,as influenced by subsequently generated typhoons in the system.These typhoons generated in the hindcast after the first 10 days attain obviously different trajectories between the two platforms.
文摘With the development of computer technology, network bandwidth and network traffic continue to increase. Considering the large data flow, it is imperative to perform inspection effectively on network packets. In order to find a solution of deep packet inspection which can appropriate to the current network environment, this paper built a deep packet inspection system based on many-core platform, and in this way, verified the feasibility to implement a deep packet inspection system under many-core platform with both high performance and low consumption. After testing and analysis of the system performance, it has been found that the deep packet inspection based on many-core platform TILE_Gx36 [1] [2] can process network traffic of which the bandwidth reaches up to 4 Gbps. To a certain extent, the performance has improved compared to most deep packet inspection system based on X86 platform at present.
文摘Simulators are generally used during the design of computer architectures. Typically, different simulators with different levels of complexity, speed and accuracy are used. However, for early design space exploration, simulators with less complexity, high simulation speed and reasonable accuracy are desired. It is also required that these simulators have a short development time and that changes in the design require less effort in the implementation in order to perform experiments and see the effects of changes in the design. These simulators are termed high-level simulators in the context of computer architecture. In this paper, we present multiple levels of abstractions in a high-level simulation of a general-purpose many-core system, where the objective of every level is to improve the accuracy in simulation without significantly affecting the complexity and simulation speed.
文摘A team of researchers from the University of Science and Technology of China(USTC)of the Chinese Academy of Sciences(CAS)and its partners have made significant advancements in random quantum circuit sampling with Zuchongzhi-3,a superconducting quantum computing prototype featuring 105 qubits and 182 couplers.
文摘目前,多核实时系统中同步任务的节能调度研究主要针对的是同构多核处理器平台,而异构多核处理器架构能够更有效地发挥系统性能。将现有的研究直接应用于异构多核系统,在保证可调度性的情况下会导致能耗变高。对此,通过使用动态电压与频率调节(Dynamic Voltage Frequency Scaling,DVFS)技术,研究异构多核实时系统中基于任务同步的节能调度问题,提出同步感知的最大能耗节省优先算法(Synchronization Aware-Largest Energy Saved First,SA-LESF)。该算法针对所有任务的速度配置进行迭代优化,直至所有任务均达到其最大限度节能的速度配置。此外,进一步提出基于动态松弛时间回收的同步感知最大能耗节省优先算法(Synchronization Aware-Largest Energy Saved First with Dynamic Reclamation,SA-LESF-DR)。该算法在保证实时任务可调度的同时,实施相应的回收策略,进一步降低系统能耗。实验结果表明,SA-LESF与SA-LESF-DR算法在能耗表现上具有优势,在相同任务集下,相比其他算法可节省高达30%的能耗。
文摘This paper presents a smart compensation system based on MCA7707 (a kind of signal processor). The li near errors and high order errors of a sensor (especially piezoresistive sensor) can be corrected by using this system. It can optimize the process of piezoresi stive sensor calibration and compensation, then, a total error factor within 0.2 % of the sensor′s repeatability errors is obtained. Data are recorded and coeff icients are determined automatically by this system, thus, the sensor compensati on is simplified greatly. For operating easily, a wizard compensation program is designed to correct every error and to get the optimum compensation.
文摘仿真点(simulation point,SimPoint)作为一种代表性采样技术被广泛应用于处理器硅前性能评估中。SimPoint为每个待评估的程序根据贝叶斯信息准则确定仿真点数目。然而,标准测试集内不同程序有着不同的行为复杂程度,需要不同数目的仿真点来准确刻画其程序行为。SimPoint无法识别出不同程序间的复杂度差异,无法做到在总仿真点数目一定的情况下,将更多的仿真点分配给行为复杂的程序以降低这些程序的性能评估误差,将更少的仿真点分配给行为简单的程序而不损失这些程序的性能评估精度。由于没有在测试集内合理地进行仿真点分配,SimPoint虽然可以给出比较准确的平均性能评估误差,但是某些行为复杂的测试子项的性能评估误差依然较大。针对这一问题,本文优化了SimPoint的仿真点局部分配方式,提出了一种全局贪心分配方法———贪心点(greedy point,GreedyPoint)方法。该方法将仿真点的分配问题抽象为含约束的优化问题,使用微架构无关特征计算表征误差,通过全局贪心算法来求解该优化问题。实验数据表明,在相同仿真开销下,与SimPoint相比,GreedyPoint可以将SPEC CPU 2017测试套件的平均性能评估误差由3.23%降低到2.08%,最大性能评估误差由21.22%大幅降低至7.01%。
