Abstract:With the increase of system scale, chip power consumption and link rate, the overall failure rate of high-performance interconnection networks will continue rising, and the traditional operation and maintenance methods will be difficult to sustain, which brings great challenges to the overall reliability and availability of HPC(high performance computing). An unsupervised algorithm prediction model for serious network failures such as network port blocking was proposed. In this model, the symptomatic rules were extracted from the history information of the switch port status register and a new feature vector was formed. The K-means clustering algorithm was used to learn and classify the feature vectors. In the prediction, the DES(double exponential smoothing) algorithm was used to predict the port state in the future through a combination of the current state of the port, and a new feature vector was obtained and K-means algorithm was used to predict whether the port blocking failure would occur. The topology information was used to build independent sub prediction models with ToR switch ports and Spine switch ports respectively, so as to further improve the accuracy of prediction. The experimental results show that the prediction model can maintain the recall rate of 88.2%, and reach the accuracy rate of 65.2%. It can provide effective early warning and guidance for the operation and maintenance personnel in the actual system.

Abstract:Predicting the job running time is beneficial to improve the scheduling performance of the system, and the clustering can help to train better prediction models. Traditional clustering algorithms are difficult to cluster similar job names. In order to better cluster similar jobs, the job name hierarchical clustering algorithm of letter-structure-number was constructed by analyzing the semantic importance of their components. Taking the real data of two supercomputers as an example, the data clustered by this algorithm was used to train the model. The experimental results show that the prediction accuracy of the model is better than that of the traditional method, and the overall prediction accuracy is 70%～80%.

Abstract:Graph coloring was used to address resource competition for the two typical computing procedures, including the flux summation and the calculation of local maximum pressure. There were three optimizations applied on graph coloring including shared memory, the reordering of volume and face indices, and the reordering of face variables. The 3D aerodynamics application with a series of mesh sizes was used in the performance test by double and single precision floating point operations on GPU Nvidia Tesla V100 and K80. The performance tests show that the shared memory is not obvious in performance. Furthermore, the reorder of volume and face indices reduces the performance of graph coloring.It is found that the reorder of face variables can increase performance remarkably. Specifically, the performance of graph coloring is increased by around 20% on V100 and 15% on K80.

Abstract:Aiming at the problem of low resource utilization of GPU (graphics processing unit) in the process of solving CFD (computational fluid dynamics), a CFD-oriented GPU resource optimization management scheme was proposed. Based on the characterization of the CFD and tasks running concurrently, a reasonable scheduling scheme was designed. By dynamically changing the startup scale and time of different tasks, our method was able to reduce resource competition while improving the effective use of GPU resources. The experimental results show that compared with the baseline method, the average speedup ratio of our proposed resource management scheme reaches 1.64x speedup under different task scales, and the use of GPU hardware resources has also been significantly improved.

Abstract:With the increasing of scale and complexity of high-performance computing systems, the mean time between failures is getting shorter, which causes an increasing probability of execution-failure caused by the hardware and software failures for parallel programs. In addition, the possible programming errors (i.e. bugs) that exist in parallel programs can also lead to execution failure. Approaches to deal with the above two types of execution failures are totally different, therefore, when an execution-failure occurs, the programmer must figure out if the failure is caused by a system fault or a programming bug. In response to this issue, a system to identifying causes of execution-failures for parallel programs was designed and implemented on the basis of the Slurm. The system has all the supported features of Slurm, as well as the ability to monitor job status, re-submit and re-run jobs. The experimental results of the job execution show that the system can distinguish the type of program execution-failures. Experiments conducted with fault injection also demonstrates the accuracy of the system.

Abstract:For the common image lossless compression methods cannot work well. Thus, a lossless compression method for synchrotron radiation source images based on image difference and neural network was proposed. The image difference method was used to reduce the linear correlations among images. The neural network was trained to learn the nonlinear correlations in the images sequence, and the pixel value was compressed with arithmetic coding using the predicted distribution. To reduce the predicting time and coding time, the pixel value was splitted into two parts for parallel compression. The tests based on the images of Shanghai Synchrotron Radiation Facility show that the proposed method can improve more than 20% in compression ratio compared to PNG(portable network graphics), JPEG2000, FLIF(free lossless image format), and the pixel value split can reduce 30% of the time in predicting and coding.

Abstract:Characteristic analysis of parallel job is the basics of workload analysis research. Job accounting log is an important data source of job characteristic analysis, however, existing tools cannot do statistics analysis with application name, because of application name not recorded in job accounting log. To solve this problem, a novel marking method for job accounting log was proposed, which based on keyword fuzzy matching. Combined with a general job data model and a flexible extensible software architecture, a parallel job feature analysis tool JobCAT was implemented. According to verification test by millions of job accounting log data from a supercomputer system, the log marking rate of JobCAT was greater than 95%. JobCAT supports 7 plugins and 29 statistical reports, and can easily make analysis report classified by application name, which has practical value to workload analysis research.

Abstract:In order to reduce the risk of failed batch jobs in co-located cloud, the K-means algorithm was used to divide batch jobs into four categories.On the basis of classification, the TLNM (two-layer nested classification model) was proposed and the prediction algorithm based on TLNM was implemented. Experiment results based on Ali Trace 2018 data set show that the ROC(receiver operating characteristic) curve of this algorithm is significantly better than other commonly used classifiers, and the area under the ROC curve (i.e.AUC) can reach 0.978, indicating that this algorithm has good classification performance. At the same time, the recall rate can reach 0.951. Through the confusion matrix, it can be seen that the TLNM algorithm can accurately predict the failed batch jobs.

Abstract:Reduction and scan are two critical primitives in parallel computing. Thus, accelerating reduction and scan shows great importance. However, the Von Neumann architecture suffers from performance and energy bottlenecks known as “memory wall” due to the unavoidable data migration. Recently, NVM (non-volatile memory) such as ReRAM (resistive random access memory), enables in-situ computing without data movement and its crossbar architecture can perform parallel GEMV (matrix-vector multiplication) operation naturally in one step. ReRAM-based architecture has demonstrated great success in many areas, e.g. accelerating machine learning and graph computing applications, etc. Parallel acceleration methods were proposed for reduction and scan primitives on ReRAM-based PIM(processing in memory) architecture, the computing process in terms of GEMV and the mapping method on the ReRAM crossbar were focused, and the co-design of software and hardware was realized to reduce power consumption and improve performance. Compared with GPU, the proposed reduction and scan algorithm achieved substantial speedup by two orders of magnitude, and the average acceleration ratio can also reach two orders of magnitude. The case of segmentation can achieve up to five (four on average) orders of magnitude. Meanwhile, the power consumption decreased by 79%.

Abstract:Aiming at the problem that improper locks in parallel programs may lead to performance bottlenecks, an approach called IdeSync was proposed to detect and optimize synchronization bottlenecks. IdeSync leveraged the static analysis to obtain the synchronized methods and blocks, and constructed a static synchronization dependency graph. The dynamic analysis technology based on the execution path was used to analyze the synchronization dependency and build the synchronization dependency graph.In order to expose the performance bottleneck, the performance change of the critical section was monitored by increasing the program workload on the synchronization dependency graph, and optimization suggestions were given for the detected synchronization bottleneck. The effectiveness of IdeSync was evaluated with 12 large real-world applications such as HSQLDB, SPECjbb2005 and RxJava, and a total of 72 synchronization bottlenecks were detected. All these bottlenecks were optimized based on IdeSync′s suggestion to achieve performance improvements, which shows that IdeSync can effectively detect and optimize synchronization bottlenecks.

Abstract:Node priority is often used to evaluate the performance of heterogeneous cluster nodes, and it is of great importance to provide suitable weight for each priority evaluation index. The AHP (analytic hierarchy process) was chosen to establish the evaluation index system of node priority, and the initial weight of each index was calculated. The BP (back propagation) neural network was then used to optimize the weights obtained by using AHP. The input of the BP neural network was the node′s performance index values collected during execution of cluster, and the output was the corresponding priority of the node. After the network training, the weight matrix was obtained and used to calculate the optimized weights. The experimental results show that the cluster node priority evaluation model based on AHP and BP can evaluate the node performance more accurately. Compared with the default resource allocation algorithm of Spark and the comparison algorithm with unoptimized weights, the cluster performance is improved effectively by using the node priority optimized. When running the same kind of load with different amount of data, the average cluster performance increases by 16.64% and 9.76%, respectively; and when running different loads with the same amount of data, the average performance of the cluster increases by 12.49% and 6.54%, respectively.

Abstract:A state-of-art deep learning method was proposed to achieve higher accuracy of MPX(multiplexing) estimation. By analyzing the similarity between the MPX results and the real data, it was proved that hardware counts gained by running the same program was linear correlated. By applying the MLP(multilayer perceptron) and Bi-GRU(bidirectional gated recurrent unit) model, the MPX data was fitted. Based on DTW (dynamic time warping), a new metric DTW-cost was proposed to judge the accuracy of MPX result. Experiment results show that when sampling 15 hardware events simultaneously, average result of 13 high performance computing applications gained by the MLP model has a 10.53% higher relative accuracy than the fixed interpolation method. The MLP model has a 19.8% improvement at most. On the hardware events which MLP has a relatively poor performance, the Bi-GRU model improved relative accuracy score by 28.8% on average.

Abstract:The existing researches on crowdsourcing task assignment don′t involve in measuring the uncertainty of workers′ qualification, and don′t achieve the collaborative assignment in many-to-many mode between tasks and workers from the angle of crowdsourcing platform. Thus, an assignment approach of software crowdsourcing tasks was proposed supporting the fuzzy measurement of workers qualification and role collaboration. Integrating the past performance of workers and the expectation of tasks′ demands, this approach employed fuzzy interval numbers to evaluate the multiple attributes ability matching degree, and aggregated the comprehensive qualification via the fuzzy analytic hierarchy process method. By introducing the role-based collaboration theory, the many-to-many software crowdsourcing tasks assignment was formulated as a combinatorial optimization problem related to a task set and a worker group, and the constraints, including the tasks′ weights, quantity of workers and potential conflicts, were used to enhance the efficiency and success rate of tasks assignment. A solution based on the CPLEX package was presented to solve the problem. Simulation experiments show that this method can efficiently and accurately realize the collaborative allocation of crowdsourcing tasks under the premise of ensuring the best completion quality of global tasks.

Abstract:In order to realize prediction of probability of successful water-exit for underwater vehicles under measured wave action and provide a rapid prediction method of underwater launch for submarine operations, the calculation model of water-exit attitude parameters of the underwater vehicle was established on the basis of the boundary element method. The water-exit attitude parameters of underwater vehicles under different wave conditions were calculated. Meanwhile, the water-exit attitude parameters under different wave conditions were integrated to form a regular wave database. The measured sea conditions in a certain period of time in the Bohai Bay were used as the launch environment, and a wave height probability distribution model conforming to the Rayleigh distribution was established. The pitch angle of the tail touching water was taken as the judging condition to realize the prediction of probability of successful water-exit for underwater vehicles under measured wave action. The calculation results show that, considering the worst conditions, the deviation of the water-exit pitch angle relative to the static water condition decreases with the increase of the wave height; under the same wave height, the deviation of the water-exit pitch angle shows a cosine variation law due to the influence of water-exit phase. The prediction method has certain reference value for submarine combat to determine the launch timing and provides a reference for the successful probability of underwater launch.

Abstract:Based on the nodal aberration theory, a special case of misalignments for reflective optical systems, which satisfying the condition of the coma-free pivot point, was studied. Under the effect of the special misalignments, the full field display aberration characteristics of the 3rd order coma and the 3rd order astigmatism were analyzed, and the analytical calculation formula of the two nodes position of the 3rd order astigmatism was established. It was found that the 3rd order coma would not change, and one node of the 3rd order astigmatism would locate near the central FOV (field of view), which revealed the drawback of using only coma-zero in axial FOV as the criterion of perfect alignment during the traditional optical assembly. Based on the analysis of wavefront aberration, a method to judge whether the system was aligned well according to the variance of wavefront aberration in only one axial FOV after introducing quantitative misalignments was proposed. A type of two-mirror optical system was adopted by CODE V(Version 10.2) to simulate the effect of misalignments on wavefront aberration, results show that the model and the method established can quantitatively analyze the two nodes position of the 3rd order astigmatism and verified the validity of the proposed judgement method for system assembly.

Abstract:In order to compare the mid infrared laser performance output by the new nonlinear crystal BaGa₄Se₇ (BGSe) and the mature commercial nonlinear crystal KTiOAsO4 (KTA), the 1.06 μm laser was used to pump BGSe (56.3 °, 0 °, type-Ⅰ) and KTA (90 °, 0 °, type-Ⅱ-A) to output 3.5 μm laser. The pump oscillation threshold of KTA and BGSe was 52.6 mJ (the theoretical value was 46.11 mJ) and 20.6 mJ (the theoretical value was 18.32 mJ) respectively when the pump wavelength was 1 064 nm, the pulse width was 13 ns, the beam diameter was 4 mm and theoptical parametric oscillator, cavity length was 90 mm. The Δλ₂/ΔT of BGSe and KTA was 3.20 nm/℃(the theoretical value was 2.49 nm/℃) and 0.073 nm/℃(the theoretical value was 0.077 nm/℃). The output linewidth of BGSe and KTA was 4.71 nm and 2.45 nm respectively. The experimental results of the pump oscillation threshold and the temperature tuning range fit the theoretical simulation well, and the results show that BGSe is better than KTA in these two aspects. But the output linewidth of KTA is narrower than that of BGSe. The experimental results indicated that BGSe is a mid-to far-infrared nonlinear crystal with wide applications.

Abstract:By making full use of the target size and shape information, a NEET (non-ellipsoidal extended target) JTC (joint tracking and classification) algorithm was proposed on the basis of the star-convex RHM (random hypersurface model). In the proposed algorithm, the target extent state was modeled as star-convex shape. By modeling the target class-related prior information with vector form, constructing its relationship with the simultaneous extent state, and integrating it into the framework of Bayesian filter, the joint processing of tracking and classification was realized. Additionally, two separate vectors were used to model the target state, and the probability update of target class was realized by particle filter based on likelihood function derivation. The simulation results show that the NEET JTC algorithm can accurately classify targets with similar size but different shapes, and improve the target state estimation results when compared with the extended target JTC algorithm based on elliptical shape. The results also show that the proposed algorithm can significantly improve the target state estimation performance when compared with the extended target tracking algorithm based on star-convex RHM.

Abstract:Collective communication is widely used in the field of high-performance computing research and engineering. In large-scale scientific and engineering computing, collective communication overhead accounts for a large proportion, sometimes even reaching 80% of the total messaging overhead. It is the performance bottleneck of the high-performance computing system. A NIC-based offloading mechanism supporting reduction operation was proposed. By embedding reduction operation logic components on NIC, the calculation of data during transmission was implemented, and the burden on the CPU and the communication delay were reduced. A 16-node protocol operation experiment was realized through the FPGA(field programmable gate array) platform, and the protocol operation in different node size was simulated based on the xNetSimPlus simulator. Experiments show that the method can effectively reduce the time of protocol operation in collective communication, and the proposed NIC offloading mechanism that supports reduction operation hardware offload can accelerate all-reduce operations by up to 2.71 times.

Abstract:The Fay-Riddell correlation, direct simulation Monte Carlo method, and the Fourier heat transfer expression based on direct simulation Monte Carlo flow field temperature were considered. Cases with different freestream Knudsen numbers (Kn) and Mach numbers (Ma) were studied to provide a new understanding of the classical continuum method to overestimate the stagnation point heat flux in the rarefied region from the microscopic perspective. The related results show that the rarefied effect on the stagnation point heat flux includes three aspects. The temperature jump upon the wall decreases the temperature gradient, reducing the stagnation point heat flux. The translational nonequilibrium was found near the wall, which leads to the failure of the Fourier heat conduction law and overestimates the heat flux. The wall-bound effect makes the Fourier heat conduction law overestimates the heat flux located three times the mean free path away from the wall.

Abstract:PC-MUSIC (phase-coherent multiple signal classification) was introduced to study the defect quantification based on the super-resolved ultrasonic image. The ultrasonic array data can be collected via full matrix capture process, and pre-processed in time domain to extract the scattered signals related with defect. The scattered signals were post-processed by PC-MUSIC to obtain the ultrasonic image with super resolution. The ultrasonic image was analyzed to extract the lateral cross section, the -6 dB main lobe width of which was defined as the assessed length of defect. The experimental system had been built, and a block of Al with a 10mm-long slot which had been machined was chosen as the tested object. It is shown that the PC-MUSIC can accurately assess the length of defect with an error less than 10% when the proper dimension of signal subspace.

Abstract:To explore the high-velocity-penetration resistant mechanism of ceramic/thin steel composite targets (hereafter called CS targets), the failure modes and penetration resistances of CS targets with 3 mm-thick SiC ceramic layer and 0.6 mm-thick steel layer were analyzed through ballistic tests, and compared with the monolithic steel plates of identical areal densities. Based on the energy conservation principle, a theoretical prediction model for high-velocity penetration of CS targets was established. Comparison between predictions and testing results was conducted. Results show that the failure modes of rear thin steel layers in CS targets change from shear plugging to petalling due to the existence of front ceramic layers, which greatly promotes the anti-penetration energy-absorption efficiency of rear thin steel layers. Therefore, the overall penetration resistances of CS targets are superior to and above 15% higher than the monolithic steel counterparts. The predicted residual velocities of projectiles after perforation of CS targets correlate well with those experimental results, and the relative errors are all within 5%, which proves the rationality and validity of present theoretical mode.

Abstract:In order to study the hydrodynamic performance of a segmented-track amphibious vehicle, and realize resistance reduction to increase speed, stern flaps were applied to the transom stern. Model towed tests and numerical simulations were carried out and both the results agreed well with each other. The longitudinal position of the center of gravity, the length and angle of the stern flap were studied and the resistance components were analyzed. Research results show that with the longitudinal position of center of gravity between 540～560 mm, the vehicle suffered least resistance. At the velocity between 3～5 m/s, the stern flap with a length of 156 mm and an included angle of 10° with the horizontal plane has the most obvious drag reduction effect. Compared with the resistance of original naked vehicle body, the resistance reduction rate is 34.3%. The installation of the stern flap can increase the hollow area at the rear of the vehicle, which is equivalent to increasing the length of waterline, thus increasing the length-to-width ratio. This research method shows that the resistance reduction and speed increase of amphibious vehicle can be effectively realized by properly adjusting the center of gravity and optimizing the parameters of wave plate.

Abstract:Based on the typical characteristics of natural ecosystems, the concept of robot ecosystem was proposed. Through the intelligent coordination and complex evolution of cluster robots, life features such as self-sustaining, self-replication and self-evolution emerged, enabling them to achieve long-term survival, reproduction and evolution under unmanned conditions, and perform specific tasks. According to the requirements of autonomous task decision-making in typical task scenarios of robot ecosystem, the characteristics of different machine learning task decision-making methods were analyzed, and the decision tree model and neural network model of autonomous task decision-making in robot ecosystem were established. The analysis shows that the accuracy of the two models is 80%~90%, and both have good stability. The results show that the autonomous task decision-making problem of robot ecosystem can be well solved by machine learning methods such as decision tree and neural network, so as to provide technical support for task application in unmanned scenes.

Abstract:In the era of digital economy, defense advanced-research management should actively face the trend of digital transformation and build a management system that is compatible with the development of the digital economy. By studying the lessons learned from civil R & D plans as well as the transformation property of digital economy, the main difficulties and requirements faced by the digital transformation of defense advanced research management were summarized and analyzed. A new paradigm for improving defense advanced-research management-digital twin defense advanced-research was put forward, in which digital twin was employed to help the digital transformation of defense advanced-research management. Furthermore, the compositions, features, capabilities, key technologies and construction stages were elaborated explicitly, in order that digital twin defense advanced-research can play a reference and promotion role for innovating the defense advanced-research management.