Quorum systems have been used to solve the problem of data consistency in distributed fault-tolerance systems. But when intrusions occur, traditional quorum systems have some disadvantages. For example, synchronous qu...Quorum systems have been used to solve the problem of data consistency in distributed fault-tolerance systems. But when intrusions occur, traditional quorum systems have some disadvantages. For example, synchronous quorum systems are subject to DOS attacks, while asynchronous quorum systems need a larger system size （at least 3f＋1 for generic data, and f fewer for self-verifying data）. In order to solve the problems above, an intrusion-tolerance quorum system （ITQS） of hybrid time model based on trust timely computing base is presented （TTCB）. The TTCB is a trust secure real-time component inside the server with a well defined interface and separated from the operation system. It is in the synchronous communication environment while the application layer in the server deals with read-write requests and executes update-copy protocols asynchronously. The architectural hybridization of synchrony and asynchrony can achieve the data consistency and availability correctly. We also build two kinds of ITQSes based on TTCB, i.e., the symmetrical and the asymmetrical TTCB quorum systems. In the performance evaluations, we show that TTCB quorum systems are of smaller size, lower load and higher availability.展开更多
在不可靠信任环境下,考虑设备到设备(Device-to-Device,D2D)辅助边缘计算中的时效性问题,针对多用户单服务器的场景,基于用户间社会关系与交互行为,构建端到端信任模型及联合服务缓存和卸载决策优化模型,以最小化平均响应时延;提出基于...在不可靠信任环境下,考虑设备到设备(Device-to-Device,D2D)辅助边缘计算中的时效性问题,针对多用户单服务器的场景,基于用户间社会关系与交互行为,构建端到端信任模型及联合服务缓存和卸载决策优化模型,以最小化平均响应时延;提出基于信任的服务缓存及卸载(Trust-based Service Caching and Task Offloading,TCO)算法,将原问题分解为多个子问题,将服务缓存子问题转换为背包问题,使用动态规划法进行求解,将D2D中继选择子问题建模为最短路径问题,使用Dijkstra算法求解,采用轮询比较方式完成最终的卸载策略。实验仿真验证了所提算法能够有效提高缓存命中率,降低用户响应时延,保障系统的时效性。展开更多
边缘侧大模型应用正成为推动智能健康、智慧城市等领域智能化与数字化进程的关键驱动力。然而,大模型海量智能任务异构性和高动态网络的不可预测性,使得边缘设备有限的算力资源难以满足复杂推理任务对高效且可靠服务质量(Quality of Ser...边缘侧大模型应用正成为推动智能健康、智慧城市等领域智能化与数字化进程的关键驱动力。然而,大模型海量智能任务异构性和高动态网络的不可预测性,使得边缘设备有限的算力资源难以满足复杂推理任务对高效且可靠服务质量(Quality of Service,QoS)的需求。因此本文提出了一种基于生成对抗网络(Generative Adversarial Network,GAN)增强的多智能体深度强化学习(Multi-Agent Deep Reinforcement Learning,MADRL)的边缘推理与异构资源协同优化方法,以实现数字孪生(Digital Twin,DT)驱动的边缘侧大模型赋能系统中异构资源的动态负载均衡,确保推理任务高效性与可靠性。首先,本文构建并分析了DT驱动的边缘侧大模型系统中的物理网络层和孪生网络层,并采用GAN实现对物理实体的孪生映射,从而对海量异构边缘数据进行分布式处理、生成与优化。接着,利用MADRL算法来对系统中的异构资源进行综合量化与协同优化,并将边缘推理数据反馈至MADRL算法中以减少集中式训练过程中的数据通信开销。同时,借助于联邦学习,该架构能够实现多方知识共享,从而有效提升模型训练速度与性能。最后,仿真结果表明,该算法能够在动态复杂大模型赋能边缘系统环境中有效降低推理任务的时延和能耗,充分利用有限的系统资源,确保推理任务的高效性,并提升智能服务的质量。展开更多
基金supported by the National Natural Science Foundation of China (60774091)
文摘Quorum systems have been used to solve the problem of data consistency in distributed fault-tolerance systems. But when intrusions occur, traditional quorum systems have some disadvantages. For example, synchronous quorum systems are subject to DOS attacks, while asynchronous quorum systems need a larger system size （at least 3f＋1 for generic data, and f fewer for self-verifying data）. In order to solve the problems above, an intrusion-tolerance quorum system （ITQS） of hybrid time model based on trust timely computing base is presented （TTCB）. The TTCB is a trust secure real-time component inside the server with a well defined interface and separated from the operation system. It is in the synchronous communication environment while the application layer in the server deals with read-write requests and executes update-copy protocols asynchronously. The architectural hybridization of synchrony and asynchrony can achieve the data consistency and availability correctly. We also build two kinds of ITQSes based on TTCB, i.e., the symmetrical and the asymmetrical TTCB quorum systems. In the performance evaluations, we show that TTCB quorum systems are of smaller size, lower load and higher availability.
文摘在不可靠信任环境下,考虑设备到设备(Device-to-Device,D2D)辅助边缘计算中的时效性问题,针对多用户单服务器的场景,基于用户间社会关系与交互行为,构建端到端信任模型及联合服务缓存和卸载决策优化模型,以最小化平均响应时延;提出基于信任的服务缓存及卸载(Trust-based Service Caching and Task Offloading,TCO)算法,将原问题分解为多个子问题,将服务缓存子问题转换为背包问题,使用动态规划法进行求解,将D2D中继选择子问题建模为最短路径问题,使用Dijkstra算法求解,采用轮询比较方式完成最终的卸载策略。实验仿真验证了所提算法能够有效提高缓存命中率,降低用户响应时延,保障系统的时效性。
文摘边缘侧大模型应用正成为推动智能健康、智慧城市等领域智能化与数字化进程的关键驱动力。然而,大模型海量智能任务异构性和高动态网络的不可预测性,使得边缘设备有限的算力资源难以满足复杂推理任务对高效且可靠服务质量(Quality of Service,QoS)的需求。因此本文提出了一种基于生成对抗网络(Generative Adversarial Network,GAN)增强的多智能体深度强化学习(Multi-Agent Deep Reinforcement Learning,MADRL)的边缘推理与异构资源协同优化方法,以实现数字孪生(Digital Twin,DT)驱动的边缘侧大模型赋能系统中异构资源的动态负载均衡,确保推理任务高效性与可靠性。首先,本文构建并分析了DT驱动的边缘侧大模型系统中的物理网络层和孪生网络层,并采用GAN实现对物理实体的孪生映射,从而对海量异构边缘数据进行分布式处理、生成与优化。接着,利用MADRL算法来对系统中的异构资源进行综合量化与协同优化,并将边缘推理数据反馈至MADRL算法中以减少集中式训练过程中的数据通信开销。同时,借助于联邦学习,该架构能够实现多方知识共享,从而有效提升模型训练速度与性能。最后,仿真结果表明,该算法能够在动态复杂大模型赋能边缘系统环境中有效降低推理任务的时延和能耗,充分利用有限的系统资源,确保推理任务的高效性,并提升智能服务的质量。
