High Performance Computing
-
consensus algorithms for distributed storage system
- SHEN Jia-jie, LU Xiu-wen, XIANG Wang, ZHAO Ze-yu, WANG Xin,
-
2022, 44(04):
571-583.
doi:
-
Abstract
(
341 )
PDF (1053KB)
(
475
)
-
Consensus algorithms are widely adopted in the distributed storage systems to ensure the correctness of the I/O operations. Since consensus algorithms typically use a complex protocol to ensure the correctness of the I/O operations between multiple storage nodes, it incurs high network transmission overhead and I/O delay. Due to the large differences in the implementation mechanisms of various consensus algorithms, specific consensus algorithms often need to be deployed in specific storage application scenarios in order to efficiently perform I/O operations and ensure the quality of service of the applications on them. Therefore, in the actual storage system development process, developers often need to select consensus algorithms according to storage application scenarios, thereby reducing the system overhead caused by I/O operations. In order to clarify the suitable application scenarios of various consensus algorithms, this paper introduces the consensus problems existing in distributed storage systems, and summarizes the implementation mechanism of current consensus algorithms. This paper summarizes the mainstream consensus algorithms under the replica-based storage systems and the erasure-coded storage systems, and compares the characteristics of these consensus algorithms in terms of implementation mechanism, network overhead, and data storage overhead. On this basis, this paper combines two classic application scenarios of a single data center distributed storage system and a cross-data center cloud-to-cloud storage system, and summarizes the main points that developers need to pay attention to when deploying consensus algorithms in actual storage systems. The problems that need to be solved urgently and the possible ways to improve the performance of I/O operations are analyzed, and the future development direction of the consensus algorithm is prospected.
-
Architecture combining active & passive performance evaluation methods and its implementation for big data storage
- LIU Shi-yuan, LI Yun-chun, CHEN Chen, YANG Hai-long
-
2022, 44(04):
584-593.
doi:
-
Abstract
(
143 )
PDF (1090KB)
(
196
)
-
Big data storage plays an important role in the whole big data application framework system in view of the increasing amount of data. Performance evaluation for big data storage can guide big data application developers to analyze performance bottlenecks and optimize the performance of big data systems. In the past, benchmarking was usually used to evaluate the performance of different big data frameworks, or the performance analysis of distributed file systems was carried out by piling and analyzing track files. These two methods adopt different analytical perspectives, but there has not been a reasonable evaluation system to evaluate the distributed storage system of big data. This paper proposes the architecture and specific implementation of the big data storage performance evaluation method combining active and passive methods. In the active evaluation method, this paper provides benchmark test programs of more than 20 applications in 6 fields to initiate performance tests on big data storage systems, and analyzes the benchmark performance indicators of big data storage systems. In the passive performance evaluation method, this paper provides analysis and positioning methods for inefficient tasks, inefficient operators and inefficient functions. By analyzing the big data applications running on the big data storage systems, we can find out the reasons for the inefficiency of big data applications. Experiments show that the architecture of the proposed big data performance evaluation method can comprehensively evaluate the performance of big data storage.
-
An application log analysis and system automation optimization framework for Lustre cluster storage
- CHENG Wen, LI Yan, ZENG Ling-fang, WANG Fang, TANG Shi-cheng, YANG Li-ping, FENG Dan, ZENG Wen-jun
-
2022, 44(04):
594-604.
doi:
-
Abstract
(
144 )
PDF (1172KB)
(
186
)
-
In the fields of scientific computing, big data processing, and artificial intelligence, it is very important to study the relevant application load, analyze the load I/O pattern to reveal the application load change law, etc., which is very important to guide the performance optimization of the cluster storage system. At present, there are many kinds of applications and the applications are updated rapidly and iteratively. The complex environment makes the feature mining of application load full of challenges. To address the above problems, we collected the application log information of five Lustre cluster storages in the production environment for 326 days, explored and analyzed the access and load characteristics of the application load, and verified and supplemented the existing observations. Through horizontal, vertical, and multi-dimensional comparative analysis and information mining of the application log information, we summarize four findings, explore the relationship between the relevant findings and previous research work, and then combine the actual production environment with the corresponding system optimization strategies. Feasible implementation schemes are given, which provide relevant references and suggestions for users, maintainers, upper application developers, multi-tier storage system designers, and other personnel. At the same time, because of the complex practical application environment and time-consuming work of system optimization, a system automation optimization framework (SAOF) is designed and implemented. SAOF can provide functions such as resource reservation and bandwidth limitation for specified application loads. Preliminary tests show that SAOF can provide automatic QoS guarantees for different tasks according to system resources and task load requirements.
-
An adaptive high-speed channel equalizer based on deep neural network
- JIAN Jie, LUO Zhang, LAI Ming-che, XIAO Li-quan, XU Wei-xia
-
2022, 44(04):
605-610.
doi:
-
Abstract
(
179 )
PDF (933KB)
(
313
)
-
High-speed serial interface is a key technology to improve the bandwidth of high-performance interconnection networks, and the channel equalizer is the core component to improve signal integrity. This paper uses the modern digital signal processing (DSP) structure to propose a deep neural network (DNN)-based high-speed channel equalization research method. This method overcomes the inherent shortcoming that the decision speed of the traditional decision feedback equalizer (DFE) is limited by the feedback loop when facing high-speed channels above 50GB in the future. The simulation results show that, compared with the traditional 2-tap DFE architecture with 15-tap FFE, the proposed 3-layer DNN structure has better equalization effect and faster equalization convergence speed, when the PAM4 encoding method is used, the high-speed channel baud rate is 28GB, and the channel loss is 15dB, or the baud rate is 56GB and the channel loss is 30dB.
-
A survey on serverless computing
- YANG Bo-ai, ZHAO Shan, LIU Fang
-
2022, 44(04):
611-619.
doi:
-
Abstract
(
266 )
PDF (723KB)
(
481
)
-
With the development of cloud computing and the transformation of traditional industries, the increasing technical requirements and booming market demands make the traditional application software development paradigm face challenges. Meanwhile, a new generation of more economical and more potential cloud service models is urgently explored. Relying on container technology, serverless comput- ing provides high concurrency and compatibility, hides the details of underlying servers, and adopts a more economical service operation model charged by the number of calls or time. So it has aroused wide concern. Firstly, the concept of serverless computing and its system architecture and technical characteristics are introduced. Then, the current research statuses of serverless computing in scientific research, open-source community and industry are introduced. Whereafter, the implementation cases of serverless computing in several application fields are listed. Last but not least, the current challenges of serverless computing technology are described.
Computer Network and Znformation Security
-
Research progress of routing techniques for data center networks
- DUAN Chen, PENG Wei, WANG Bao-sheng
-
2022, 44(04):
631-644.
doi:
-
Abstract
(
147 )
PDF (676KB)
(
216
)
-
Due to the rapid development of cloud computing, the demand of Internet service providers for data centers is increasing rapidly. As a key technology of data center networks, routing is responsible for selecting routes for network traffic within or across data centers. It can also provide diffe- rential services for traffic with different QoS (Quality of Service) requirements. However, for large-scale data centers, due to the unpredictable traffic of applications and the topology characteristic of data centers, the traditional routing methods in the Internet cannot provide satisfactory high throughput and resource utilization. Network congestion make it hard to provide guaranteed bandwidth or transmission delay for flows with the demand of QoS. This paper classifies and compares the routing problems in data centers, and then emphatically introduces the research progress about the unicast routing methods. Further, the congestion-aware routing methods are introduced. Finally, the routing technology of the new data center networks is discussed.
-
A DV-Hop positioning algorithm combining sine and cosine optimization and hop distance optimization
- ZHANG Jing , HE Yuan-yuan,
-
2022, 44(04):
645-653.
doi:
-
Abstract
(
97 )
PDF (642KB)
(
136
)
-
In order to solve the problems of imprecise calculation of the hop distance and inaccurate positioning due to the inability of the least squares solution to reach the optimal unbiased state in the DV-Hop positioning algorithm, a DV-Hop positioning algorithm combining sine and cosine optimization and hop distances optimization is proposed, which defines the concept of the optimised anchor node. Firstly, the anchor node with the smallest average hop distance among all anchor nodes around each unknown node is selected as the optimized anchor node in the algorithm, then any other anchor node is selected to form a triangle with other unknown nodes, and the edge from the optimal anchor node to the unknown node is considered to be the optimal edge in the triangle. Secondly, the distances from the other anchor nodes to the unknown nodes are calculated to optimize the hop distances by the law of cosine. Finally, the sine-cosine optimization algorithm is used to improve the least square method, and the volatility of the sine-cosine function is used to find the optimal position of the unknown node. The experimental results indicate that, compared with the conventional DV-Hop and DV-Hop improvement algorithms, the proposal reduces the positioning error significantly.
-
A mobile proxy application traffic identification method based on machine learning
- CUI Hong, ZHAO Shuang, ZHANG Guang-sheng, SU Jin-shu
-
2022, 44(04):
654-664.
doi:
-
Abstract
(
265 )
PDF (1173KB)
(
290
)
-
With the rapid development of mobile networks, more users choose to protect privacy, hide online behavior and bypass the restrictions of networks by using proxy applications. As a result, new challenges are brought to network management and auditing. In addition, malicious attackers can use proxy to hide their identity, making it more difficult to detect and prevent such malicious behavior. Therefore, proxy application traffic identification plays an important role in network management and security, while this issue has not been fully studied at present. Because the proxy application traffic is usually encrypted and obfuscated, the traditional traffic identification methods can not be applied effectively. To achieve accurate and fast traffic identification of mobile proxy applications, a set of side- channel traffic features that are independent of the payload is proposed. The option field in the TCP header is used for the first time to describe the traffic characteristics. Four machine learning algorithms with two kinds of identification objects are utilized to validate the effectiveness and importance of the proposed feature set. The experimental results show that the proposed features can effectively identify proxy application traffic. More than 99% accuracy can be achieved when identifying whether traffic is forwarded by proxy applications based on random forest. Moreover, the average accuracy is higher than 94% when identifying which proxy application the traffic belongs to. Compared with other methods, the proposed method has better accuracy and faster classification speed on the public dataset ISCX VPN- nonVPN. Hence, it is more suitable for real-time traffic identification scenarios.
-
Optimization of dynamic feature selection algorithm for malicious behavior detection
- LIU Yun, XIAO Tian, WANG Zi-yu
-
2022, 44(04):
665-673.
doi:
-
Abstract
(
116 )
PDF (875KB)
(
127
)
-
For malicious behaviors existing in the Internet, especially online malicious user behavior detection in social network applications, clustering analysis algorithms based on multi-dimensional user characteristics are usually used for detection. This paper proposes a dynamic feature selection algorithm (DFSA), which uses a fuzzy C-means objective function with feature weighted entropy. Firstly, a learning mode is constructed for the parameters, and each feature weight is automatically calculated, and features whose weight is less than the threshold are eliminated. Important feature components are selected dynamically, and the membership function, cluster center and feature weights are updated iteratively until the optimization is achieved. Finally, malicious user behavior clusters with high accuracy is detect- ed. The simulation results show that the proposed algorithm outperforms the SDAFS algorithm, the ELAFC algorithm and the NADMB algorithm in terms of three main performance indicators such as Rand index, Jaccard index and normalized mutual information.
-
Advances in deep learning methods for pavement crack detection and identification with visual images
- LU Kai-liang
-
2022, 44(04):
674-685.
doi:
-
Abstract
(
324 )
PDF (2662KB)
(
455
)
-
Surface crack identification with visual images is a kind of non-contact detection solution, which is not limited by the material of the tested object and is easy to achieve online automation. Therefore, it has the advantages of fast speed, low cost and high precision. Firstly, the public data sets of typical pavement crack are comprehensively collected, and the characteristics of sample images and the random variable factors are summarized. Subsequently, the advantages and disadvantages of hand-crafted feature engineering, machine learning, and deep learning crack identification methods are compared. Finally, from the aspects of network architecture, testing performance and predicting effectiveness, this paper reviews the development and progress of typical deep learning algorithms such as self-built CNN, transfer learning (TL) and encoder-decoder (ED) that can be easily trained and deployed. We can see the obvious improvement of performance and effect because of algorithm optimization and computing power enhancement. The results show that real-time patch-level and fast pixel-level crack detection can be realized on low computing power GPU platform.
-
An improved interactive multi-model tracking algorithm for maneuvering targets
- YANG Dong-ying, HE Jiang-peng
-
2022, 44(04):
686-691.
doi:
-
Abstract
(
110 )
PDF (778KB)
(
171
)
-
In order to improve the tracking accuracy of maneuvering targets and obtain more accurate real-time position and velocity information of the targets, an improved interactive multi-model tracking algorithm is proposed. Based on the traditional interactive multi-model, this algorithm introduces adjustment parameters associated with target features. The target feature data is used to provide a limited domain for the initial data, and then the adjustment parameters are added to the filter, so as to complete the optimization of the tracking accuracy by the iteration of the target state gain matrix and the covariance matrix. Experimental simulation analyzes three common states of maneuvering targets and compares them with traditional interactive multi-model tracking algorithms. The experimental results show that, when the deflection angles of the adjustment parameters of the algorithm are 4°, 2° and 1°, the mean root mean square errors are 15.91 m, 11.79 m and 11.39 m respectively, which are significantly better than the traditional algorithms mean value of 21.39 m. At the same time, with the improvement of parameter setting accuracy, it has a certain inhibitory effect on the error fluctuation caused by the maneuvering target state.
-
A train bottom parts detection algorithm based on OSE-dResnet neural networks
- LI Li-rong, WANG Zi-yan, ZHANG Kai, YANG Di-chun, XIONG Wei, GONG Peng-cheng,
-
2022, 44(04):
692-698.
doi:
-
Abstract
(
139 )
PDF (777KB)
(
142
)
-
Aiming at the difficulty of the train bottom parts detection, a detection algorithm based on OSE-dResnet network is proposed. In order to increase the richness and accuracy of feature extraction, feature extraction is enhanced by increasing cross-layer transmission, based on the Resnet50 network. Second, an OSEnet module is embedded into the feature extraction network to enhance beneficial feature channels with global features. Finally, the feature layers of different scales are fused to achieve the feature information complementation. The experimental results show that the algorithm combining the OSEnet module and the d-Resnet network has a good detection effect on the bottom parts of the train. The proposed algorithm is validated on the test datasets, and mAP reaches 98.77%.
Artificial Intelligence and Data Mining
-
SNM algorithm optimization based on field filtering and scaling window
- ZHOU Shi-jie, LOU Yuan-sheng
-
2022, 44(04):
699-706.
doi:
-
Abstract
(
108 )
PDF (1050KB)
(
163
)
-
The problematic data in the data warehouse has a great impact on data quality. In order to find and delete these problematic data, the primary work is the processing of similar repeated data. Currently, the most widely used algorithm for deduplication is the sorted-neighborhood method (SNM). After analyzing the shortcomings of this algorithm, an improved SNM algorithm (ISNM) is proposed. The attribute weights are calculated using the attribute discrimination method, which solves the subjectivity caused by artificial weights. The field filtering algorithm is used to calculate the similarity of two records, which reduces the number of comparisons of record attributes in the window and accelerates the detection speed of the algorithm. Variable windows are used instead of fixed-size windows to prevent missing records and reduce useless record comparisons. Experimental results show that ISNM algorithm has obvious advantages in terms of recall, precision and running time overhead.
-
Comparison and analysis of ED algorithm and SNP-index algorithm in calculating SNP sites——Take arabidopsis thaliana for example
- GAN Qiu-yun
-
2022, 44(04):
707-712.
doi:
-
Abstract
(
187 )
PDF (806KB)
(
168
)
-
SNP (Single Nucleotide Polymorphism) is the most common variation in biological heritable variation, which occurs between single nucleoside acid-base groups in DNA sequence. ED algorithm and SNP-index algorithm are two commonly used algorithms to calculate SNP sites. The whole genome sequencing data of F2 generation of arabidopsis thaliana are obtained by high-throughput sequencing. The sequencing data are filtered, screened and compared based on Linux platform. The number of SNP sites and the proportion of SNP genotypes detected under different algorithms are compared. The experimental results show that the number of SNP sites obtained by ED algorithm is more and more widely distributed than SNP index algorithm, and the relative distribution density is larger than that of SNP index algorithm, but the number of SNP sites and the proportion of SNP genotypes obtained by the two algorithms are similar.
-
A rational label propagation algorithm based on node influence
- HUANGFU Fei-fei, YANG Yang, DENG Xiao-yi,
-
2022, 44(04):
713-722.
doi:
-
Abstract
(
152 )
PDF (2559KB)
(
278
)
-
Community discovery can reveal the topology and important nodes of real social networks. Due to its linear time complexity and no need to define objective functions and objective parameters, Label Propagation Algorithm (LPA) is widely used in academic and practical fields as a classic community discovery algorithm. Aiming at the update disorder of LPA algorithm and the randomness of label selection, a Rational Node Label Propagation Algorithm Based on Node Influence (RLPBNI) is proposed. The algorithm takes the node influence ranking as the update order, introduces the concept of rational nodes for label selection, and defines the overlap degree for community dimensionality reduction. The analysis of the experimental results shows that, compared with other comparative algorithms, the RLPBNI algorithm can not only effectively improve the accuracy of community division, but also more easily discover hidden communities in networks with a high degree of mixing.
-
Feature selection based on general importance and runner-root algorithm
- WU Shang-zhi, XU Dan-dan, WANG Xu-wen, XIA Ning
-
2022, 44(04):
723-729.
doi:
-
Abstract
(
113 )
PDF (594KB)
(
126
)
-
Feature selection is an important step in the data preprocessing stage in machine learning, pattern recognition, data mining and other fields. In reality, the data information collected is of high dimension, and there are redundant data and noisy data, which will increase the calculation time and mislead the modeling results at the same time. Combined with the generalized importance of attribute subsets and the intelligent optimization runner-root algorithm, a feature selection algorithm is proposed. The method uses the runner-root algorithm for iterative optimization, and uses the generalized importance of attribute subsets and the size of the selected feature subsets as fitness functions to evaluate the selected feature subsets, so that the features that are important for decision making are searched out as far as possible in the entire sample space. The experimental results show that the proposed feature selection algorithm can select effective feature subsets and obtain higher accuracy on the classification model.
-
Unsupervised neural machine translation model based on pre-training
- XUE Qing-tian, LI Jun-hui, GONG Zheng-xian, XU Dong-qin
-
2022, 44(04):
730-736.
doi:
-
Abstract
(
146 )
PDF (655KB)
(
185
)
-
Depending on the large-scale parallel corpus, neural machine translation has achieved great success in some language pairs. Subsequently, unsupervised neural machine translation (UNMT) has partly solved the problem that high quality corpus is difficult to obtain. Recent studies show that cross-lingual language model pretraining can significantly improve the translation performance of UNMT. This method models deep context information in cross-lingual language scenarios by using a large-scale monolingual corpus, and obtains significant results. This paper further explores UNMT based on cross-lingual language pretraining, proposes several improved methods of training model, and compares the performance between UNMT and baseline system on different language pairs. Aiming at the issue of unbalanced initialization of unsupervised NMT parameters when using pre-trained models, this paper proposes a secondary pre-training stage to continue pre-training, and propose to initialize the Cross attention sub-layer with the self-attention sub-layer in unsupervised NMT model. Meanwhile, as back- translation plays a critical role in unsupervised NMT, we propose to use Teacher-Student framework to guide back-translation.Experimental results show that, compared with the baseline system, these methods improve BLEU by 0.8~2.08 percentages at most.
-
Efficient semantic segmentation based on improved DeepLabV3+
- MA Dong-mei, LI Peng-hui, HUANG Xin-yue, ZHANG Qian, YANG Xin
-
2022, 44(04):
737-745.
doi:
-
Abstract
(
436 )
PDF (1267KB)
(
342
)
-
The current high-precision semantic segmentation medel generally have the problems of high computational complexity and large memory usage, so it is difficult to deploy on embedded platforms with limited hardware storage and computing power. Aiming at the problem, an improved efficient semantic segmentation medel based on improved DeepLabV3+ is proposed by comprehensively considering three aspects of network parameters, calculation and performance. The model uses MobileNetV2 as the backbone network, and combines the mix strip pooling(MSP) in the atrous spatial pyramid pooling (ASPP) module to obtain dense context information. The effective channel attention (ECA) module is introduced in the decoder to restore a clearer target boundary. Depthwise separable convolution is applied to the ASPP module and decoder to compress the model. Experiment on the PASCAL VOC 2012 dataset show that the number of network parameters of the medel is 4.5×106, the number of floating point operations is 11.13 GFLOPs, and the mean intersection over union is 72.07%, which proves that the algorithm achieves the good balance between calculation efficiency and segmentation accuracy.
-
Path planning of artificial potential field method based on simulated annealing algorithm
- ZHAO Bing-wei, JIA Feng, CAO Yan, SUN Yu, LIU Yi-hong
-
2022, 44(04):
746-752.
doi:
-
Abstract
(
255 )
PDF (1134KB)
(
463
)
-
In the traditional artificial potential field method, the local minimum point problem exists in path planning, which makes the mobile robot unable to move to the target point. Therefore, an artificial potential field method based on simulated annealing algorithm is proposed. Artificial potential field method uses the simulated annealing algorithm to add random target points near the local minimum point and guide the mobile robot to escape from the local minimum point area gradually. Finally, Matlab simulation proves that the method can make the mobile robot escape from the local minimum position, successfully reach the target position, consumes shorter time, and is more stable.
