Serving multiple inference tasks of deep neural networks (DNNs) concurrently on a shared GPU is an established method for maximizing hardware resource.Although DNN compilers effectively generate optimal kernel code for individual DNN inferences, they fall short in optimizing for concurrent tasks.This paper presents ConCo, a concurrencyaware compilation scheme designed to optimize the execution of concurrent DNN inference tasks on a shared GPU.ConCo dynamically generates multiple code variants, each tailored to different GPU resource constraints, and efficiently selects optimal variants at runtime according to concurrent workload characteristics.To mitigate the substantial overhead associated with multi-variant compilation, ConCo employs an optimal-code-sharing strategy, significantly accelerating compilation by leveraging commonalities across resource configurations.Evaluations demonstrate that ConCo improves inference throughput by up to 1.2× and reduces job completion time by up to 69.85% compared to existing solutions.

Insufficient dynamic adaptability and lagging response to sparse events in sluice gate scheduling due to sudden rainfall in urban water-sheds is a critical topic, which raises significant importance and practical interest. This study proposes a scheduling framework based on Deep Q-Network (DQN) integrated with Support Vector Regression (SVR) and enhanced by Prioritized Experience Replay (PER). By modeling sluice gate scheduling as a sequential decision-making problem within a Markov Decision Process (MDP) model, the scheduling strategy is effectively optimized using Reinforcement Learning (RL). To address the modeling challenges of traditional DQN in high-dimensional state spaces, we incorporate a SVR model to predict the trend of water level changes. In addition, the PER is introduced to dynamically adjust the priority of experience replay, with a focus on improving the learning efficiency for sparse high-risk events. Experiments on the Nanjing Wuding men sluice gate show that the proposed scheduling framework achieves a maximum reduction of approximately 90.9 % in dispatch frequency under one of the three water level control criteria, while maintaining stable water level control across all scenarios.

Accurate flood prediction is crucial for disaster prevention and mitigation. Hydrological data exhibit highly nonlinear temporal patterns and encompass complex spatial relationships between rainfall and flow. Existing flood prediction models struggle to capture these intricate temporal features and spatial dependencies. This paper presents an adaptive periodic and spatial self-attention method based on LSTM (APS-LSTM) to address these challenges. The APS-LSTM learns temporal features from a multi-periodicity perspective and captures diverse spatial dependencies from different period divisions. The APS-LSTM consists of three main stages, (i) Multi-Period Division, that utilizes Fast Fourier Transform (FFT) to divide various periodic patterns; (ii) Spatio-Temporal Information Extraction, that performs periodic and spatial self-attention focusing on intra- and inter-periodic temporal patterns and spatial dependencies; (iii) Adaptive Aggregation, that relies on amplitude strength to aggregate the computational results from each periodic division. The abundant experiments on two real-world datasets demonstrate the superiority of APS-LSTM. The code is available: https://github.com/oopcmd/APS-LSTM.

The topic models can effectively reduce the dimensionality of document representation. In recent years, Generative Adversarial Networks (GANs) have begun to be applied to topic models, but most researches do not fully consider the documents’ context information, leading to the fact that the generated results of the encoder networks cannot represent the real data distribution. To this end, we propose a novel topic model named Semantic-based Bidirectional Adversarial Neural Topic Model (SNTM), which introduces semantic information into Bidirectional Generative Adversarial Networks (BiGAN) by adding the word embedding and BiLSTM-Attention mechanism. This improvement makes the encoder network in our model generate a distribution closer to the real document-topic distribution. Furthermore, we also propose a topic difference evaluation indicator, which more comprehensively evaluates the quality of the generated topics. The experimental results on different datasets show that SNTM performs better than the baseline topic model.

Inferring causal relationships is key to data science. Learning causal structures in the form of directed acyclic graphs (DAGs) has been widely adopted for uncovering causal relationships, nonetheless, it is a challenging task owing to its exponential search space. A recent approach formulates the structure learning problem as a continuous constrained optimization task that aims to learn causal relation matrix. Following it are various variants, such as nonlinear variants which uncover nonlinear causal relationships. However, the nonlinear variant which considers an ordinary sparsity constraint as part of its optimization objective may not effectively eliminate spurious edges, i.e., wrongly predicted edges. In this paper, we investigate the defect that the ordinary sparsity constraint inherently contains, that is, when the model predicts spurious edges, the sparsity constraint may be unable to make the relation matrix sparse to help eliminate such spurious edges, and based on the theoretical and empirical analysis of the defect, we propose the proportional sparse penalty constraint PSP which replaces the ordinary sparsity constraint with a normalized first-order matrix norm. We then compare our proposed model GAE-PSP with three models to show considerable performance improvement. We further conduct three comparative experiments and show that the PSP can make the relation matrix sparse to eliminate spurious edges.

The intelligent question answering over the knowledge graph aims to automatically answer natural language questions via locating the correct entities in the knowledge graph. Aside from the former progresses, it is still challenging to answer the multi-relation questions because of the variety and complexity of the natural language, as well as the combinatorial explosion on possible candidates. In this paper, we propose a novel embedding-based approach named SPE-QA to address these issues. It answers a question by identifying its most semantic like question-answer path from the candidate topic-entity-centric subgraph, and locating this path's tail entity as the final answer. In order to limit the scale of the candidate sub graph and thus reduce the neural network's training complexity, it is essential to filter the explicit noises as much as possible. Therefore, we employ a sub graph separation method to decide the scale of the sub graph, and remove the false question-answer paths with two different pruning policies. The experimental results on two widely used benchmarks approve that, our mechanism can not only reduce the subgraph's scale dramatically, but also obtain better performance on the multi-relation questions, compare to the stat-of-the-art approaches.

Efficient scheduling algorithms are key for attaining high performance in heterogeneous computing systems. In this article, we propose a new list scheduling algorithm for assigning task graphs to fully connected heterogeneous processors with an aim to minimize the scheduling length. The proposed algorithm, called Improved Predict Priority Task Scheduling (IPPTS) algorithm has two phases: task prioritization phase, which gives priority to tasks, and processor selection phase, which selects a processor for a task. The IPPTS algorithm has a quadratic time complexity as the related algorithms for the same goal, that is$O(t^{2} \times p)$, for$t$tasks and$p$processors. Our algorithm reduces the scheduling length significantly by looking ahead in both task prioritization phase and processor selection phase. In this way, the algorithm is looking ahead to schedule a task and its heaviest successor task to the optimistic processor, i.e., the processor that minimizes their computation and communication costs. The experiments based on both randomly generated graphs and graphs of real-world applications show that the IPPTS algorithm significantly outperforms previous list scheduling algorithms in terms of makespan, speedup, makespan standard deviation, efficiency, and frequency of best results.

As the scale of RDF data increases gradually, the requirements for efficient processing technologies on RDF data in the distributed environment has become a hot topic. Most existing methods have two main issues. First, massive storage redundancy often occurs to materialize semi-join results, but only a part is frequently accessed in most queries. Second, the queries are allocated to the whole cluster, causing a considerable duplication on intermediate results and hence increasing the communication cost. In this paper, we propose a new approach named FP-ExtVP, using a hybrid partitioning schema to reduce the storage requirement without damaging the query performance. In addition, we create an index named P-tree to allocate the queries only to the mentioned nodes, instead of the whole cluster, in order to reduce the shuffle cost. The experiments on several benchmarks demonstrate that, comparing to the native S2RDF, our method outperforms on Star, Snowflake and Complex queries, by reducing nearly a half of the storage requirements.

Efficient workflow scheduling algorithms are crucial for attaining high performance in large-scale heterogeneous distributed infrastructures, such as cloud computing. List scheduling algorithms are one of the most efficient heuristic methods for assigning task graphs to fully connected heterogeneous systems. However, most existing list-based scheduling algorithms do not consider the applications' security requirements and the security services offered by cloud providers. In this paper, we extend four list scheduling algorithms for security-aware workflow scheduling in the IaaS cloud. The idea of the extension is to consider the security overheads in both tasks prioritizing phase and virtual machine selection phase of the four original algorithms. Based on real-world applications, we evaluate the performance of the proposed algorithms in terms of scheduling length, speedup and efficiency.

Machine learning models, particularly reinforcement learning (RL), have demonstrated great potential in optimizing video streaming applications. However, the state-of-the-art solutions are limited to an "offline learning" paradigm, i.e., the RL models are trained in simulators and then are operated in real networks. As a result, they inevitably suffer from the simulation-to-reality gap, showing far less satisfactory performance under real conditions compared with simulated environment. In this work, we close the gap by proposing OnRL, an online RL framework for real-time mobile video telephony. OnRL puts many individual RL agents directly into the video telephony system, which make video bitrate decisions in real-time and evolve their models over time. OnRL then aggregates these agents to form a high-level RL model that can help each individual to react to unseen network conditions. Moreover, OnRL incorporates novel mechanisms to handle the adverse impacts of inherent video traffic dynamics, and to eliminate risks of quality degradation caused by the RL model's exploration attempts. We implement OnRL on a mainstream operational video telephony system, Alibaba Taobao-live. In a month-long evaluation with 543 hours of video sessions from 151 real-world mobile users, OnRL outperforms the prior algorithms significantly, reducing video stalling rate by 14.22% while maintaining similar video quality.

With the recent advancements in the information and communication technologies, large number of devices are connecting to the Internet, hence large volumes of data in different formats and from different sources are generating. Consequently, on one hand dynamic and heterogeneous data sharing and management, in the ecosystem of Internet of Things (IoT), where every smart object is connected to Internet, presents new research challenges. On the other hand, citizen privacy preserving is another challenge, because he/she has to send his/her information to a service provider, to obtain the required information. This information is sensitive since it can reveal information about an individual. An attacker or a malicious service provider can utilize this sensitive information for their own business or something else. This paper presents a Secure Framework for Future Smart City (SEFSCITY), for better city living and governance, based on Cloud Computing IoT and Distributed Computing. We first present the architecture of SEFSCITY, which is based on Multi-Cloud and Cloud Federation approach; then we propose a security protocol for our framework. In our security model, we use Zero-Knowledge Protocol based on Elliptic Curve Discrete Logarithm Problem. Finally, we validate our architecture by conducting several scenarios that we have implemented using Cloud Analyst tool. The results show that in all scenarios, the cost infrastructure remains the same for the cloud customer, and our approach is benefic for the cloud provider in term of revenues and data processing time

Traditionally the data retrieval is achieved by searching the metadata with keywords, though it is often difficult for ordinary users to express professional and precise query demands in the water industry. Regarding this issue, this paper introduces an exploratory retrieval method called faceted search by gradually recommending relevant to the users. Firstly, a unified modeling algorithm is proposed to construct the unified metadata model in XML for heterogeneous water metadata. Based on this model, candidate facet terms can be extracted and filtered, in order to retrieve the various water metadata uniformly. At last, a facet recommendation algorithm is proposed to help the users to sharpen their queries by prompting less but more accurate as the search gets deeper, after excluding those irrelevant facets and redundant facets. The experimental results demonstrate that our facet recommendation algorithm can significantly improve the retrieval precision on querying the water data.

The spatial join is usually processed in two steps: filter and refinement. The first creates candidate pairs based on the objects' abstractions, like object pairs having overlapping minimum bounding rectangles (MBRs). The second step then further checks each candidate pair whether the objects fit the join condition with their actual shapes. This two-step strategy prevents the disk I/O and CPU overhead spent on retrieving and comparing the non-candidates, hence it is also adopted at the large scale processing, based on the Hadoop platform. However, in order to cope with different data distributions, the Hadoop-based spatial join often processes both above steps in the Reduce stage, causing considerable data migration overhead as all object pairs need to be shuffled over the cluster network. In this paper we propose two novel shuffling approaches for the distributed filter and refinement operations. They are able to retrieve and even transfer only the candidates' actual shapes data in the Reduce tasks, decreasing the migration overhead to the minimum. In our evaluations with a six-node cluster, both approaches outstandingly improve the process efficiency regardless of the join selectivity.

Parallel Secondo scales up the capability of processing extensible data models in Secondo. It combines Hadoop with a set of Secondo databases, providing almost all existing SECONDO data types and operators. Therefore it is possible for the user to convert large-scale sequential queries to parallel queries without learning the Map/Reduce programming details. This paper demonstrates such a procedure. It imports the data from the project OpenStreetMap into Secondo databases to build up the urban traffic network and then processes network-based queries like map-matching and symbolic trajectory pattern matching. All involved queries were stated as sequential expressions and time-consuming in single-computer Secondo. However, they can achieve an impressive performance in Parallel Secondo after being converted to the corresponding parallel queries, even on a small cluster consisting of six low-end computers.

This paper presents a hybrid parallel processing system, named Parallel Secondo. It combines the Hadoop framework and a set of single-computer Secondo databases, in order to introduce the mobility data procedures into the parallel processing community, and vice versa. The system keeps the front-end and the executable language of Secondo to allow the users to state their parallel queries like common sequential queries. Besides, a set of auxiliary scripts is provided so as to make it easier to manage the system no matter how large the underlying cluster is, and keep the Hadoop platform as a transparent level of the system. Further, a parallel data model is also proposed in this paper to encapsulate all available Secondo data types and operators. Thereby, it is able to transform any Secondo sequential query to its corresponding parallel expression. For instance, all example queries in the moving objects database benchmark BerlinMOD are transformed, and two of them are demonstrated in this paper. In the last evaluations, this paper illustrates that Parallel Secondo is not only a practical but also an efficient system. For queries involving large amounts of data, it performs both linear speed-up and scale-up.

The growing need of processing massive amounts of data leads database researchers to explore the possibility of combining their existing single-computer database systems with the popular parallel processing platform Hadoop. These hybrid systems not only can keep the efficiency of database processing, but also achieve a remarkable scalability.

Hadoop is an efficient and simple parallel framework following the Map Reduce paradigm, and making the parallel processing recently become a hot issue in data-intensive applications. Since Hadoop can be easily deployed on large-scale clusters including up to thousands of computers, various studies intend to process common relational database operations also on this new platform and expect to achieve a remarkable performance. However, these works have to prepare customized programs according to different input format, making the communication between co-workers difficult. Additionally, all intermediate data have to be transformed to key-value pairs and then transferred through the underlying HDFS, making the data processable by Map and Reduce tasks and keeping a balanced workload on the cluster. During this period, unnecessary overhead decreases both the speed-up and scale-up of these systems. Therefore, this paper attempts to propose a light and efficient coupling structure thus to combine Hadoop with single-computer databases on the engine level. On one hand, it uses a well-designed parallel data model to make end-users represent parallel queries like common queries. All current and future data types and algorithms can be used directly, having no need to be specifically changed for the parallel platform. On the other hand, it provides a simple and independent distributed file system to transfer data among database engines directly, without passing through HDFS, hence to remove as much as possible unnecessary transform and transfer overhead. For purpose of demonstration, a prototype Parallel Secondo is introduced in this paper. It has been fully evaluated in both small and large scale clusters, achieving satisfactory performances for different database operations.