检索结果-内蒙古大学图书馆

20th ifip WG 10.3 international conference on network and parallel computing, NPC 2024

作者： Wang, Huiyu Liu, Zhicheng Wang, Xiaofei Qiu, Chao Zhang, Cheng Wang, Wenyu Ye, Qianwen College of Intelligent and Computing Tianjin University Tianjin China Tianjin University of Finance and Economics Tianjin China Shanghai Zhuichu Networking Technologies Co. Ltd. Shanghai China Hubei Central China Technology Development of Electric Power Co. Ltd. Wuhan China

ISBN: (纸本)9789819628636

In recent years, the rapid rise of Artificial Intelligence Generated Content (AIGC) has been accompanied by a surge in demand for generative AI application services. this trend has not only driven the diversified deployment of large-scale models but also posed more significant challenges to the efficient inference of these models. Current computing solutions have struggled to solely keep pace with growing user demands, which results in insufferable latency and inefficient performance. Integrating cloud computing and edge computing for the inference process of large models is deemed a promising solution, while related resource scheduling issues are still pending. To address this challenge, we focus our research on diffusion models, a representative model in the field of generative artificial intelligence, and propose a cloud-edge collaborative framework to accelerate the inference pipeline of text-to-image diffusion models. Our optimization goal is to minimize the average completion time of inference tasks while maintaining image quality. Considering that the inference for each image is an iterative process, we propose a step clipping strategy to save and reuse intermediate results to reduce iterations, compressing the calculation amount of inference tasks. Above this, we design a task scheduling strategy using proximal policy optimization, which can make efficient resource decisions. Experiments demonstrate the superiority of our strategy over local execution, greedy task scheduling strategy, and learning-based task scheduling strategies. Compared with the three baselines, the average task completion time under different system configurations is reduced by 1.093 s, 1.024 s, and 0.481 s, respectively. © ifip international Federation for Information Processing 2025.

关键词： Cloud platforms

来源：评论

学校读者我要写书评

暂无评论

8th international Symposium on Combinatorial Optimization, ISCO 2024

8th International Symposium on Combinatorial Optimization, I...

引用

8th international Symposium on Combinatorial Optimization, ISCO 2024

ISBN: (纸本)9783031609237

the proceedings contain 30 papers. the special focus in this conference is on Combinatorial Optimization. the topics include: Quadratically Constrained Reformulation, Strong Semidefinite Programming Bounds, and Algorithms for the Chordless Cycle Problem;a Family of Spanning-Tree Formulations for the Maximum Cut Problem;optimal Cycle Selections: An Experimental Assessment of Integer Programming Formulations;1-Persistency of the Clique Relaxation of the Stable Set Polytope;alternating Direction Method and Deep Learning for Discrete Control with Storage;branch and Cut for Partitioning a Graph into a Cycle of Clusters;computing the Edge Expansion of a Graph Using Semidefinite Programming;minimizing External Vertices in Hypergraph Orientations;open-Separating Dominating Codes in Graphs;on the Complexity of the Minimum Chromatic Violation Problem;crystal Trees;Reducing Treewidth for SAT-Related Problems Using Simple Liftings;total Matching and Subdeterminants;a New Structural Parameter on Single Machine Scheduling with Release Dates and Deadlines;fixed-Parameter Algorithms for Cardinality-Constrained Graph Partitioning Problems on Sparse Graphs;sequencing Stochastic Jobs with a Single Sample;the thief Orienteering Problem on Series-parallel Graphs;approximation Algorithm for Job Scheduling with Reconfigurable Resources;network Design on Undirected Series-parallel Graphs;online Graph Coloring with Predictions;neuron Pairs in Binarized Neural networks Robustness Verification via Integer Linear Programming;optimal Counterfactual Explanations for k-Nearest Neighbors Using Mathematical Optimization and Constraint Programming;surrogate Constraints for Synchronized Energy Production/Consumption;a Robust Two-Stage Model for the Urban Air Mobility Flight Scheduling Problem;optimal Charging Station Location in a Linear Cycle Path with Deviations;an Efficient Timing Algorithm for Drivers with Rest Periods;fair Energy Allocation for Collective Self-consumption;day-Ahead Lot-Sizing Un

关键词：

来源：评论

学校读者我要写书评

暂无评论

Online harmonic extraction and synchronization algorithm based control for unified power quality conditioner for microgrid systems

引用

ENERGY REPORTS 2022年 8卷 962-971页

作者： Hasan, Kamrul Othman, Muhammad Murtadha Meraj, Sheikh Tanzim Rahman, Nor Farahaida Abdul Noor, Siti Zaliha Mohammad Musirin, Ismail Abidin, Izham Zainal Univ Teknol MARA Solar Res Inst SRI Sch Elect Engn Coll Engn Studies Shah Alam 40450 Selangor Malaysia Univ Teknol PETRONAS Dept Elect & Elect Engn Seri Iskandar 32610 Perak Malaysia Univ Tenaga Nas Coll Engn Dept Elect Power Engn Jalan IKRAM UNITEN Kajang 43000 Selangor Malaysia

this manuscript proposes an Adaptive Linear Neural network (ADALINE) based control for the unified power quality conditioner to protect the critical loads from voltage related power quality problems and mitigate the current related power quality problems produced by nonlinear loads. Due to parallel computing feature and simple algorithm, ADALINE is utilized widely in parameter estimation. In this manuscript, unified ADALINEs based current harmonic extractor for the shunt active power filter is utilized, and an ADALINE based control technique is utilized to produce reference voltages for load voltage compensation. Due to ADALINE control, the necessity of a phase locked loop and conventional filters are omitted, therefore, enabling the system to extract the fundamental components more accurately. Besides, the ability of self-regulating transition from islanded mode to grid-connected mode has ensured that uninterrupted power supply is provided to the critical loads. the system performance is validated through simulation results utilizing MATLAB-Simulink software considering several dynamic conditions, namely unbalanced grid voltage, self-regulating transition, and grid unavailability. (c) 2021 the Author(s). Published by Elsevier Ltd. this is an open access article under the CC BY license (http://***/licenses/by/4.0/). Peer-review under responsibility of the scientific committee of the 8th international conference on Power and Energy Systems Engineering, CPESE 2021.

关键词： Battery energy storage Microgrid Neural network Power quality Solar photovoltaic Unified power quality conditioner

来源：评论

学校读者我要写书评

暂无评论

FNotify: A Low-Latency and Scalable Publish/Subscribe System using RDMA 24

FNotify: A Low-Latency and Scalable Publish/Subscribe System...

引用

24th IEEE international conference on High Performance computing and Communications, 8th IEEE international conference on Data Science and Systems, 20th IEEE international conference on Smart City and 8th IEEE international conference on Dependability in Sensor, Cloud and Big Data Systems and Application, HPCC/DSS/SmartCity/DependSys 2022

作者： Xu, Yusen Shen, Sijie Han, Mingcong Chen, Rong Institute of Parallel and Distributed Systems Seiee Shanghai Jiao Tong University Shanghai China

ISBN: (纸本)9798350319934

Publish/subscribe systems are widely used as messaging middleware in various scenarios, such as online business, financial trading, and social media. However, the network overhead intro-duced by the pub/sub system can severely impact application performance. Especially in the latency-sensitive applications of the data center, optimizing the latency of publishing or delivering messages of the pub/sub system is particularly important. RDMA network has gained popularity in relevant applications due to its low latency, high bandwidth and ability to bypass the CPU. Never-theless, it is still challenging to design and optimize the pub/sub system using RDMA. through an in-depth analysis of RDMA features, we design and implement FNotify, a high-performance RDMA-accelerated pub/sub system. FNotify optimizes latency and CPU load using in-memory message management and one-sided RDMA primitives for message collection. FNotify also proposes a topic-based partitioning scheme to achieve good scal-ability. We conducted a comprehensive evaluation and analysis on an 8-node RDMA cluster to verify the validity of our design by comparing the performance of FNotify with Apache Kafka (a state-of-the-art pub/sub system) in a typical pub/sub application scenario and various micro-benchmarks. Experimental results show that FNotify improves overall performance by 1.5× to 3.9× and reduces end-to-end latency by more than 20× compared to Apache Kafka. © 2022 IEEE.

关键词： Middleware

来源：评论

学校读者我要写书评

暂无评论

LBoDSN: An In-network Load Balancing Mechanism for Lossless Data Center networks Based on Direct Switch Notification 20th

LBoDSN: An In-Network Load Balancing Mechanism for Lossless...

引用

20th ifip WG 10.3 international conference on network and parallel computing, NPC 2024

作者： Shi, Qingyu Jiang, Fangxue Huang, Huang Li, Xiaocui Li, Chuang Cao, Wenzhi Liu, Limei School of Advanced Interdisciplinary Studies Hunan University of Technology and Business Changsha China Xiangjiang Laboratory Changsha China

ISBN: (纸本)9789819628292

With the extensive deployment of Remote Direct Memory Access (RDMA) in lossless data center networks (DCNs), enhancing the RDMA load balancing performance for distributed AI training and HPC applications becomes particularly critical. However, existing load balancing schemes either suffer difficulty in responding to congestion in sub-RTT time, or cannot accurately detect and reroute flows that are on the verge of creating path congestion for Priority-based Flow Control (PFC) enabled lossless DCNs. In this paper, we propose LBoDSN, an in-network load balancing mechanism for lossless DCNs based on direct swtich notification, to address above challenges. LBoDSN monitors ingress queue length evolution at destination switches to predict the triggering time of PFC pause, and accurately identifies congested flows based on the congestion contribution level before PFC pause, then further proactively sends the flow congestion notification (FCN) to source switches for fast rerouting. the FCN contains a flow ID that identifies the congested flow, and a path ID that identifies the target path to which the congested flow will be rerouted. And after rerouting, the Congestion Notification Packet (CNP) of the old path is selectively discarded at source switches to improve transmission performance, while out-of-order packets are reordered at the destination switch. the experimental results show that under realistic workloads, LBoDSN achieves 9–64% and 20–80% better than CONGA for the average and tail Flow Completion Times (FCTs), respectively. Compared with ConWeave, LBoDSN achieves around 8% better performance for the average and tail FCTs while significantly reducing switch queue usage for reordering. © ifip international Federation for Information Processing 2025.

关键词： Time switches

来源：评论

学校读者我要写书评

暂无评论

Power Estimation Models for Edge computing Devices

Power Estimation Models for Edge Computing Devices

引用

29th international conference on parallel and Distributed computing (Euro-Par)

作者： Kasioulis, Michalis Symeonides, Moysis Pallis, George Dikaiakos, Marios D. Univ Cyprus Dept Comp Sci Nicosia Cyprus

ISBN: (纸本)9783031506833;9783031506840

the increasing demand for energy-efficient solutions in IoT devices and edge computing calls for novel methodologies to generate accurate power models for diverse devices, enabling sustainable growth and optimized performance. this paper presents a methodology for creating power models for edge devices and their embedded components. the proposed methodology collects power and resource utilization measurements from the edge device and generates both additive and regression models. the methodology is evaluated on a Raspberry Pi 4 device using a smart plug for power monitoring and various benchmarking tools for CPU and network sub-components. the evaluation shows that the generated models achieve low error, demonstrating the effectiveness of the proposed approach. Our methodology can be applied to any edge device, providing insights into the most efficient power consumption model. the heterogeneity of edge devices poses a challenge to creating a global power model, and our approach provides a solution for developing device-specific power models. Our results indicate that the generated models for Raspberry Pi 4 scored a maximum of 8% MAPE.

关键词： Power Consumption Power Modeling IoT Edge computing Edge Benchmarking

来源：评论

学校读者我要写书评

暂无评论

PimCity: A Compute in Memory Substrate featuring both Row and Column parallel computing 8

PimCity: A Compute in Memory Substrate featuring both Row an...

引用

8th IEEE international conference on Rebooting computing, ICRC 2023

作者： Resch, Salonik Cilasun, Husrev Zabihi, Masoud Chowdhury, Zamshed Zhao, Zhengyang Wang, Jian-Ping Sapatnekar, Sachin S. Karpuzcu, Ulya University of Minnesota Twin Cities United States

ISBN: (纸本)9798350382044

Processing-in-memory (PIM) substrates can perform parallel computation directly within the memory array. As a result, throughput performance and energy efficiency can reach unprecedented levels, however, there are limitations. Typical PIM architectures only support parallel computing in one dimension of the memory array: Computation is performed along either multiple rows or multiple columns (but not both). this restricts data layout and makes intra-array intermediate data transfers during computation inevitable - which require reads and writes, even for short-distance data movement. Inter-array data transfers, on the other hand, become a problem for larger scale algorithms which use more than one PIM array. Such data transfers incur large communication overheads and increase the complexity of the peripheral circuitry and interconnection network between arrays. Intermediate data transfers of any form limit scalability and efficiency. To overcome this limitation, we introduce PimCity, a new PIM substrate which can compute in both the rows and the columns of the memory array. PimCity replaces intraarray data transfer (memory) operations with lower overhead logic operations inside the memory array. Further, PimCity can perform logic operations directly across neighboring memory arrays, which in turn enables low-cost inter-array data transfers. PimCity hence represents a scalable PIM architecture suitable for both HPC and embedded IoT style applications. © 2023 IEEE.

关键词： Data transfer

来源：评论

学校读者我要写书评

暂无评论

MiDedup: A Restore-Friendly Deduplication Method on Docker Image Storage Systems 20th

MiDedup: A Restore-Friendly Deduplication Method on Docker ...

引用

20th ifip WG 10.3 international conference on network and parallel computing, NPC 2024

作者： Qin, Lisha Tan, Haoliang Zou, Xiangyu Ou, Wenhao Wei, Yan Huang, Rubing Xia, Wen Harbin Institute of Technology Shenzhen China Peng Cheng Laboratory Shenzhen China Zhongruan International Technology Service Co. Ltd. Shenzhen China Macau University of Science and Technology Macao China

ISBN: (纸本)9789819628292

Docker image storage systems, like Docker registry, often employ deduplication to reduce storage overhead. Existing deduplication methods for these systems detect redundancy at either layer or file level, each with its own advantages and limitations. Layer-level deduplication offers much faster restoration but lower deduplication ratio, as it eliminates duplicate layers while ignoring duplicate files across layers. Conversely, file-level deduplication offers a higher deduplication ratio but slower restoration, as removing duplicate files scatters the files of a layer, leading to multiple disk read I/O during restoration. Our analysis reveals that the slow restoration in file-level deduplication partly stems from ignoring file-layer relationships. In this paper, we propose MiDedup, a new Docker deduplication method. MiDedup detects redundancy at file-level and organizes files into a restore-friendly layout by considering file-layer relationships to enhance restore speed. Our studies also show that duplicated files of one layer are not uniformly distributed across different layers. MiDedup uses this insight to skip deduplicating scattered files to further optimize the file layout for faster restoration. Evaluations indicate that MiDedup achieves restore speeds that are 8–19× faster than file-level deduplication (slightly slower than layer-level one), while maintaining 9–47% higher deduplication ratios than layer-level deduplication. © ifip international Federation for Information Processing 2025.

关键词： Redundancy

来源：评论

学校读者我要写书评

暂无评论

Adaptive Splitting Algorithm for Neural network Modelling in AMP 21

Adaptive Splitting Algorithm for Neural Network Modelling in...

引用

21st IEEE international Symposium on parallel and Distributed Processing with Applications, 13th IEEE international conference on Big Data and Cloud computing, 16th IEEE international conference on Social computing and networking and 13th international conference on Sustainable computing and Communications, ISPA/BDCloud/SocialCom/SustainCom 2023

作者： Zhang, Kaifei Ma, Yuxiang Wu, Yulei Chen, Tao Ma, Huijie Henan University School of Computer and Information Engineering Kaifeng475004 China University of Bristol School of Electrical Electronic and Mechanical Engineering BristolBS8 1UB United Kingdom

ISBN: (纸本)9798350329223

Artificial Intelligence algorithms usually consume a lot of energy and computing resources during training, so data is usually sent to high-performance devices to complete model training. However, such a computing mode cannot make good use of computing resources in the network and is not conducive to the privacy protection of user information. In recent years, some studies have used end devices to participate in model training, among which federated learning is a typical representative. Due to the heterogeneous computing resources of the end devices, how to split the neural network model to achieve an effective balance between training time and energy consumption is an important issue. In this paper, we propose a novel model splitting method that includes the steps of collecting data, correcting the data, and determining the splitting layers to complete the splitting learning. We consider four factors (training time, energy consumption, CPU utilisation, and memory utilisation) to jointly decide the splitting layer and propose a scheme for adaptively revising the model. We use the VGG16 model for evaluation. the results show that the efficiency of our method is improved by 64.5% compared to the recently proposed ARES scheme. At the same time, our proposed method does not affect the convergence and accuracy of the model. © 2023 IEEE.

关键词： Neural network models

来源：评论

学校读者我要写书评

暂无评论

CH-Go: Online Go System Based on Chunk Data Storage 24

CH-Go: Online Go System Based on Chunk Data Storage

引用

作者： Lu, Hui Li, Chuan Yang, Yiming Li, Cheng Islam, Ashraful Computer School Sichuan University Chengdu China School of Computing National University of Singapore Singapore Singapore

ISBN: (纸本)9798350319934

the training and running of an online Go system require the support of effective data management systems to deal with vast data, such as the initial Go game records, the feature data set obtained by representation learning, the experience data set of self-play, the randomly sampled Monte Carlo tree, and so on. Previous work has rarely mentioned this problem, but the ability and efficiency of data management systems determine the accuracy and speed of the Go system. To tackle this issue, we propose an online Go game system based on the chunk data storage method (CH-Go), which processes the format of 160k Go game data released by Kiseido Go Server (KGS) and designs a Go encoder with 11 planes, a parallel processor and generator for better memory performance. Specifically, we store the data in chunks, take the chunk size of 1024 as a batch, and save the features and labels of each chunk as binary files. then a small set of data is randomly sampled each time for the neural network training, which is accessed by batch through yield method. the training part of the prototype includes three modules: supervised learning module, reinforcement learning module, and an online module. Firstly, we apply Zobrist-guided hash coding to speed up the Go board construction. then we train a supervised learning policy network to initialize the self-play for generation of experience data with 160k Go game data released by KGS. Finally, we conduct reinforcement learning based on REINFORCE algorithm. Experiments show that the training accuracy of CH-Go in the sampled 150 games is 99.14%, and the accuracy in the test set is as high as 98.82 %. Under the condition of limited local computing power and time, we have achieved a better level of intelligence. Given the current situation that classical systems such as GOLAXY are not free and open, CH-Go has realized and maintained complete Internet openness. © 2022 IEEE.

关键词： Reinforcement learning

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：