基于NR-Transformer的集群作业运行时间预测

高性能集群的作业调度通常使用作业调度系统来实现，准确填写作业运行时间能在很大程度上提升作业调度效率。现有的研究通常使用机器学习的预测方式，在预测精度和实用性上还存在一定的提升空间。为了进一步提高集群作业运行时间预测的准确率，考虑先对集群作业日志进行聚类，将作业类别信息添加到作业特征中，再使用基于注意力机制的NR-Transformer网络对作业日志数据建模和预测。在数据处理上，根据与预测目标的相关性、特征的完整性和数据的有效性，从历史日志数据集中筛选出7维特征，并按作业运行时间的长度将其划分为多个作业集，再对各作业集分别进行训练和预测。实验结果表明，相比于传统机器学习和BP神经网络，时序神经网络结构有更好的预测性能，其中NR-Transformer在各作业集上都有较好的性能。     

单片机应用系统综合实验装置的设计

单片机原理与应用作为电子技术类学科中的必修课程具有很强实践性,要求学生掌握实际应用能力,为此研制了基于单片机系统通信为基础的实验板,依据单片机应用系统特点,可完成SPI、I^2C、UART串行通信以及并行总线方式通信。实验板有利于学生更好地了解和掌握单片机应用系统设计的方法,有利于学生进行自主设计和内容拓展。做到培养学生的学习能力、提高学生的学习兴趣、发挥学生的创造性。     

High Accuracy Network Cardinalities Estimation by Step Sampling Revision on GPU

Host cardinality estimation is an important research field in network management and network security. The host cardinality estimation algorithm based on the linear estimator array is a common method. Existing algorithms do not take memory footprint into account when selecting the number of estimators used by each host. This paper analyzes the relationship between memory occupancy and estimation accuracy and compares the effects of different parameters on algorithm accuracy. The cardinality estimating algorithm is a kind of random algorithm, and there is a deviation between the estimated results and the actual cardinalities. The deviation is affected by some systematical factors, such as the random parameters inherent in linear estimator and the random functions used to map a host to different linear estimators. These random factors cannot be reduced by merging multiple estimators, and existing algorithms cannot remove the deviation caused by such factors. In this paper, we regard the estimation deviation as a random variable and proposed a sampling method, recorded as the linear estimator array step sampling algorithm (L2S), to reduce the influence of the random deviation. L2S improves the accuracy of the estimated cardinalities by evaluating and remove the expected value of random deviation. The cardinality estimation algorithm based on the estimator array is a computationally intensive algorithm, which takes a lot of time when processing high-speed network data in a serial environment. To solve this problem, a method is proposed to port the cardinality estimating algorithm based on the estimator array to the Graphics Processing Unit (GPU). Experiments on real-world highspeed network traffic show that L2S can reduce the absolute bias by more than 22% on average, and the extra time is less than 61 milliseconds on average.     

有限元SN中子输运模拟的区域分解并行

确定论中子输运方法具有计算速度快、可获取物理量的精细场分布、可高效多物理耦合等优点，随着有限元方法在中子输运模拟中的应用，复杂几何结构、大尺度下的屏蔽问题和临界问题都能得到高保真建模和分析。离散纵标（S_N）法是求解中子输运方程的有效数值方法，基于OpenMP并行机制对各独立离散方向进行并行求解，可提高S_N输运模拟的计算速度，但并行规模较有限。对几何空间进行区域分解并采用MPI并行机制，可实现大规模并行扩展，进而实现对大型问题的高精度快速求解。本文采用并行自适应非结构网格应用框架JAUMIN进行区域分解和进程间通信，通过并行S_N扫描实现了自主有限元输运程序ENTER的高效并行，完成正确性检验后在天河Ⅱ号超级计算机上使用1 440个CPU核完成了1.43×10⁷网格单元、2.81×10⁹自由度规模问题的测试，计算时间约7.4 h。表明该程序具备了有效模拟大型复杂结构中子输运问题的能力，具有一定工程应用价值。     

Mixed coordination mechanisms for scheduling games on hierarchical machines

In this paper, we study scheduling games under mixed coordination mechanisms on hierarchical machines. The two scheduling policies involved are ‐ and ‐, where ‐ (resp., ‐) policy sequences jobs in nondecreasing order of their hierarchies, and jobs of the same hierarchy in nonincreasing (resp., nondecreasing) order of their processing times. We first show the existence of a Nash equilibrium. Then we present the price of anarchy and the price of stability for the games with social costs of minimizing the makespan and maximizing the minimum machine load. All the bounds given in this paper are tight.     

基于双向位图的CSR大规模图存储优化

大数据时代,Graph500是评测超级计算机处理数据密集型应用能力的重要工具,E级验证系统的图遍历处理能力主要受限于内存空间和访存带宽,尤其是内存空间利用率直接决定了图的测试规模和测试性能.针对天河E级验证系统小内存特征,提出了基于双向位图的大规模图数据压缩存储方法(bidirectional-bitmap based CSR,Bi-CSR),Bi-CSR在CSR矩阵压缩的基础上引入行方向位图和列方向位图协同完成稀疏矩阵压缩存储,行方向位图主要负责行方向位图的压缩存储与索引,列方向位图除了进一步压缩图存储空间,还负责为顶点遍历向量并行优化提供加速空间.Bi-CSR大幅度减少了稀疏矩阵存储空间.面向天河E级验证系统,当图输入规模为237时,Graph500的图存储空间节约效率接近70%,全系统稳定测试性能为2.131E+12TEPS,性能最大加速比超过100倍.     

格点QCD基础求解器及其异构计算实现的性能优化

格点量子色动力学(格点QCD)是研究夸克、胶子等微观粒子间相互作用的重要理论和方法. 通过将时空离散化为四维结构网格, 并将量子色动力学的基本场量定义在网格上, 让研究人员可以使用数值模拟方法, 从第一性原理出发研究强子间相互作用和性质, 但这个过程中的计算量极大, 需要进行大规模并行计算. 格点QCD计算的核心基础为格点QCD求解器, 是程序运行主要的计算热点模块. 本文研究在国产异构计算平台下格点QCD求解器的实现与优化, 提出一套格点QCD求解器的设计实现, 实现了BiCGSTAB求解器, 显著降低了迭代次数; 通过对奇偶预处理技术, 降低了所求问题的计算规模; 针对国产异构加速卡的特点, 优化了Dslash模块的访存操作. 实验测试表明, 相比优化前的求解器获得了约30倍的加速比, 为国产异构超算下格点QCD软件性能优化提供了有益的参考价值.     

Designing amorphous formulations of polyphenols with naringin by spray-drying for enhanced solubility and permeability

Naringin (NAR), a major flavanone (FVA) glycoside, is a component of food mainly obtained from grapefruit. We used NAR as a food additive to improve the solubility and permeability of hydrophobic polyphenols used as supplements in the food industry. The spray-dried particles (SDPs) of NAR alone show an amorphous state with a glass transition temperature (T_g) at 93.2 °C. SDPs of hydrophobic polyphenols, such as flavone (FVO), quercetin (QCT), naringenin (NRG), and resveratrol (RVT) were prepared by adding varying amounts of NAR. All SDPs of hydrophobic polyphenols with added NAR were in an amorphous state with a single T_g, but SDPs of hydrophobic polyphenols without added NAR showed diffraction peaks derived from each crystal. The SDPs with NAR could keep an amorphous state after storage at a high humidity condition for one month, except for SDPs of RVT/NAR. SDPs with NAR enhanced the solubility of hydrophobic polyphenols, especially NRG solubility, which was enhanced more than 9 times compared to NRG crystal. The enhanced solubility resulted in the increased membrane permeability of NRG. The antioxidant effect of the hydrophobic NRG was also enhanced by the synergetic effect of NAR. The findings demonstrated that NAR could be used as a food additive to enhance the solubility and membrane permeability of hydrophobic polyphenols.     

分段中深孔采矿法切割槽及底部出矿结构设计优化

间柱采场用分段中深孔嗣后充填采矿法回采，为利用中深孔爆破形成切割槽及底部桃形体出矿结构，优化切割天井位置，以溜矿口上方的桃形体作为安全隔柱，利用溜矿口空间作为底部拉槽的凿岩和装药空间；采用平行中深孔和扇形中深孔相结合的方式进行布眼，爆破形成切割槽和底部桃形体出矿结构，避免了采用人工浅孔爆破施工的安全风险,切割槽爆破到位，底部桃形体出矿结构成型良好且稳固安全，该技术已在全矿推广应用。     

The provably good parallel seeding algorithms for the k-means problem with penalties

As a classic NP-hard problem in machine learning and computational geometry, the k-means problem aims to partition the given points into k sets to minimize the within-cluster sum of squares. The k-means problem with penalties (k-MPWP), as a generalizing problem of the k-means problem, allows a point that can be either clustered or penalized with some positive cost. In this paper, we mainly apply the parallel seeding algorithm to the k-MPWP, and show sufficient analysis to bound the expected solution quality in the case where both the number of iterations and the number of points sampled in each iteration can be given arbitrarily. The approximate guarantee can be obtained as , where is a polynomial function involving the maximal ratio M between the penalties. On one hand, this result can be viewed as a further improvement on the parallel algorithm for k-MPWP given by Li et al., where the number of iterations depends on other factors. On the other hand, our result also generalizes the one solving the k-means problem presented by Bachem et al., because k-MPWP is a variant of the k-means problem. Moreover, we present a numerical experiment to show the effectiveness of the parallel algorithm for k-means with penalties.