Search

After the service provider temporarily selects the required edge nodes based on social and storage capabilities, application execution causes the edge nodes to cache part of the application data. Therefore, offloading part of the application computing tasks to the selected edge nodes can effectively improve application execution performance. However, in cases where the resources of user's IoT devices are insufficient, tasks can be further offloaded to traditional edge servers or even to the cloud to maximize application execution efficiency. In this paper, the entire uninstall utility is modeled as a weighted sum of task completion time and energy consumption. Under the premise of considering users' preferences for completion time and energy consumption, a game-based uninstallation algorithm is proposed. The algorithm performs uninstallation by optimizing the uninstallation decision. Based on user preferences, the total system overhead is relatively small. The subsequent simulation experiments show that the algorithm can reduce system overhead on the basis of satisfying user preferences and has relatively good adaptability. [ABSTRACT FROM AUTHOR]

With the rapid development of cloud computing technology and the broad application of the Internet, more and more user data are stored in the cloud storage system online. Data copy technology is an effective technical means to manage various data in cloud storage systems, which has the advantages of improving data access speed and enhancing data availability in cloud storage systems. Since a data copy management strategy has an important impact on the performance of cloud storage systems, how to design a data copy management strategy is of great significance to achieving high-quality cloud storage management. Firstly, this paper analyzes the HDFS random copy placement strategy and its shortcomings. Secondly, proposes a copy placement strategy based on evaluation value and load balancing. Thirdly, gives the evaluation function of rack and node by considering multiple evaluation indicators such as node disk space load capacity, node memory utilization, node CPU utilization, rack load rate and network distance between racks. Fourthly, finds suitable nodes through computer shelf evaluation values and node evaluation values when placing replicas. Finally, through simulation experiments, the experimental results of the HDFS random replica placement strategy and the evaluation results and load balancing-based replica placement strategy are compared and analyzed. [ABSTRACT FROM AUTHOR]

Cloud computing platform is a data-centric architecture, which utilize virtualization resource scheduling technology to integrate and distribute the processing power of distributed server cluster and other network devices distributed over the network, and uses service-oriented architecture to provide users with highly reliable, highly available, and efficient application data services. Although the cloud platform application based on various requirements has become a research hotspot in cloud computing technology, in some high Real-time requirements application scenarios, the research on how to guarantee the real time performance of cloud platform is few. Compared to traditional resource virtualization such as CPU and memory, Docker only provides a small portion of network resource virtualization. The four network modes provided by Docker officially have a single function, which makes it very difficult to apply to cloud platform scenarios with complex network bandwidth requirements and high real-time requirements. Based on Docker's traditional NAT communication method, this paper proposes a network virtualization solution based on SDN and Docker containers. Through comparison experiments, it is proved that the scheme can guarantee the Real-time performance of the container cloud platform. [ABSTRACT FROM AUTHOR]

As a fast-rising storage model in recent years, cloud storage has adopted a "pay-as-you-go" approach to provide users with highly reliable, highly available, low-cost, and secure storage services that have received widespread attention and use within enterprises and individuals. Data replica management technology, as an essential part of cloud storage systems, has irreplaceable advantages in improving cluster fault tolerance and availability, so it has become the focus of many experts and scholars. It replicates multiple data blocks and places them in various nodes in the cluster; this makes the data more secure and reliable and improves the access rate while ensuring system load balance. Data replica technology runs through the process from replica creation to consistency maintenance. Each part of it has an essential impact on the performance of the cloud storage system. This article focuses on the dynamic decision of the number of data replicas in the cloud storage system. Considering the shortcomings of the static replica strategy, a dynamic decision strategy for the number of replicas based on the popularity of the data is proposed. By using the gray prediction model GM (1, 1) to predict the future data access frequency and using the Markov model to modify the prediction result, the data can be divided into hot data and non-hot data according to the predicted value, thereby determining the data replica number. Finally, through simulation experiments, the experimental results of the static replica strategy and the data hot-based replica number dynamic decision strategy are compared and analyzed. [ABSTRACT FROM AUTHOR]

The application of data deduplication technology reduces the demand for data storage and improves resource utilization. Compared with limited storage capacity and computing capacity of a single node, cluster data deduplication technology has great advantages. However, the cluster data duplication technology also brings new issues on deduplication rate reduction and load balancing of storage nodes. The application of data routing strategy can well balance the problem of deduplication rate and load balancing. The paper introducesa global BloomFilter routing strategy. In order to avoid the communication overhead caused by sending the fingerprints to the data storage node and inquiring about the BloomFilter maintained in the memory, a BloomFilter array is maintained in the memory of the client-server. Each row of the array corresponds to a storage node; before sending the Superchunk to the storage node, it will inquire the BloomFilter array and storage capacity information to get the optimal node. The theoretical analysis and experimental results prove the feasibility of the strategies proposed by this paper. [ABSTRACT FROM AUTHOR]

The optimization of traditional FTL(Flash Translation Layer) algorithm is mainly aimed at the average response time of flash memory read and write operation on a flash chip, because of the Out-of-place update, traditional FTL algorithm writes data in a new free page each time, and a new block is assigned when a block is full. Therefore, when the flash memory is almost full, a written request will lead to a garbage collection, in which there will be many write copies, which will lead to a significant decrease in response time. In this paper, we optimize the traditional FTL algorithm and propose an FTL algorithm that guarantees real-time performance by shortening the worst response time of the request as a measure of the guaranteed flash Qos (Quality of Service) under the condition that the address mapping is provided the upper layer. The algorithm adopts the idea of space-for-time and spreads the written copy of flash memory to each write request so that the worst response time of the write request is significantly reduced. The algorithm also wears balancing considerations to the algorithm. After experimental verification, the proposed algorithm has advantages over existing FTL algorithms in terms of real-time performance and has a performance improvement of more than 11% in terms of average response time and wear balance. By the way, we add the wear-level module into an algorithm. According to improving the algorithm, the worst response time has shorten 47.6%. [ABSTRACT FROM AUTHOR]

As user-generated data diversifies, an increasing number of big data tasks now encompass unstructured data forms, including audio, imagery, and video content. The processing of these data often requires the support of GPU devices in the cluster. However, most of the existing cluster resource management frameworks lack effective decoupling of CPU resources and GPU resources, and the management granularity of GPU resources is too coarse, resulting in poor GPU sharing and low resource utilization. Given, for the inadequacy of the existing cluster resource management framework that cannot support batch stream integration and CPU-GPU resource scheduling, we design and implement a Hybrid Heterogeneous Resource Management (H-HRM) for CPU-GPU heterogeneous clusters. The resource queue binding mechanism on computing nodes provides flexible binding of CPU resources to GPU resources, which solves the problem that CPU resources and GPU resources in the cluster are difficult to decouple. According to the different uses of CPU or GPU as the main resource, a Hybrid Domain Resource Fairness (HDRF) model is proposed to realize the reasonable allocation of CPU resources and GPU resources. Through the queue stacking technology and the mechanism that multiple executors can run simultaneously on the queue, fine-grained sharing of GPU resources is realized and the utilization rate of GPU resources is improved. Finally, this paper realizes the docking of H-HRM and Spark programming framework and conducts real load tests, and after comparing with the performance of Mesos, it is proved that H-HRM can handle the mixed scenarios of batch jobs and stream processing jobs in the cluster, and the HDRF algorithm also greatly improves the utilization of the GPU. [ABSTRACT FROM AUTHOR]

This paper mainly focuses on the I/O cache of the server layer device of the network block device in the embedded background. It proposes a method focusing on different read/write loading features, configuring the cache capacity ratio utilized for isolated processing reading data and writing data, and an algorithm for automatically analyzing sequential addressing writing and random addressing writing data streams. The effectiveness of the algorithm for automatically analyzing sequential addressing writing and random addressing writing data streams is verified by experiments. The experimental results show that according to the different interleaving conditions of the data stream, the proportion of the data stream can be written in sequential addressing, which can be different in all input data streams. After separating sequential data streams interleaved by three inputs, the percentage of data streams written in sequential addressing is significantly increased by 50%. [ABSTRACT FROM AUTHOR]

Deduplication is a popular data reduction technology in storage systems which has significant advantages, such as finding and eliminating duplicate data, reducing data storage capacity required, increasing resource utilization, and saving storage costs. The file features are a key factor that is used to calculate the similarity between files, but the similarity calculated by the single feature has some limitations especially for the similar files. The storage node feature reflects the load condition of the node, which is the key factor to be considered in the data routing. This paper introduces a multifeature data routing strategy (DRMF). The routing strategy is made based on the features of the cluster, including routing communication, file similarity calculation, and the determination of the target node. The mutual information exchange is achieved by routing communication, routing servers, and storage nodes. The storage node calculates the similarity between the files stored, and then the file is routed according to the information provided by the routing server. The routing server determines the target node of the route according to the similar results and the node load features. The system prototype is designed and implemented; also, we develop a system to process the feature of cluster and determine the specific parameters of various features of experiments. In the end, we simulate the multifeature data routing and single-feature data routing, respectively, and compare the deduplication rate and data slope between the two strategies. The experimental results show that the proposed data routing strategy using multiple features can improve the deduplication rate of the cluster and maintain a lower data skew rate compared with the single-feature-based routing strategy MCS; DRMF can improve the deduplication rate of the cluster and maintain a lower data skew rate. [ABSTRACT FROM AUTHOR]