As an important complement to cloud computing, edge computing can effectively reduce the workload of the backbone network. To reduce latency and energy consumption of edge computing, deep learning is used to learn the...As an important complement to cloud computing, edge computing can effectively reduce the workload of the backbone network. To reduce latency and energy consumption of edge computing, deep learning is used to learn the task offloading strategies by interacting with the entities. In actual application scenarios, users of edge computing are always changing dynamically. However, the existing task offloading strategies cannot be applied to such dynamic scenarios. To solve this problem, we propose a novel dynamic task offloading framework for distributed edge computing, leveraging the potential of meta-reinforcement learning (MRL). Our approach formulates a multi-objective optimization problem aimed at minimizing both delay and energy consumption. We model the task offloading strategy using a directed acyclic graph (DAG). Furthermore, we propose a distributed edge computing adaptive task offloading algorithm rooted in MRL. This algorithm integrates multiple Markov decision processes (MDP) with a sequence-to-sequence (seq2seq) network, enabling it to learn and adapt task offloading strategies responsively across diverse network environments. To achieve joint optimization of delay and energy consumption, we incorporate the non-dominated sorting genetic algorithm II (NSGA-II) into our framework. Simulation results demonstrate the superiority of our proposed solution, achieving a 21% reduction in time delay and a 19% decrease in energy consumption compared to alternative task offloading schemes. Moreover, our scheme exhibits remarkable adaptability, responding swiftly to changes in various network environments.展开更多
In a cloud environment,graphics processing units(GPUs)are the primary devices used for high-performance computation.They exploit flexible resource utilization,a key advantage of cloud environments.Multiple users share...In a cloud environment,graphics processing units(GPUs)are the primary devices used for high-performance computation.They exploit flexible resource utilization,a key advantage of cloud environments.Multiple users share GPUs,which serve as coprocessors of central processing units(CPUs)and are activated only if tasks demand GPU computation.In a container environment,where resources can be shared among multiple users,GPU utilization can be increased by minimizing idle time because the tasks of many users run on a single GPU.However,unlike CPUs and memory,GPUs cannot logically multiplex their resources.Additionally,GPU memory does not support over-utilization:when it runs out,tasks will fail.Therefore,it is necessary to regulate the order of execution of concurrently running GPU tasks to avoid such task failures and to ensure equitable GPU sharing among users.In this paper,we propose a GPU task execution order management technique that controls GPU usage via time-based containers.The technique seeks to ensure equal GPU time among users in a container environment to prevent task failures.In the meantime,we use a deferred processing method to prevent GPU memory shortages when GPU tasks are executed simultaneously and to determine the execution order based on the GPU usage time.As the order of GPU tasks cannot be externally adjusted arbitrarily once the task commences,the GPU task is indirectly paused by pausing the container.In addition,as container pause/unpause status is based on the information about the available GPU memory capacity,overuse of GPU memory can be prevented at the source.As a result,the strategy can prevent task failure and the GPU tasks can be experimentally processed in appropriate order.展开更多
基金funded by the Fundamental Research Funds for the Central Universities(J2023-024,J2023-027).
文摘As an important complement to cloud computing, edge computing can effectively reduce the workload of the backbone network. To reduce latency and energy consumption of edge computing, deep learning is used to learn the task offloading strategies by interacting with the entities. In actual application scenarios, users of edge computing are always changing dynamically. However, the existing task offloading strategies cannot be applied to such dynamic scenarios. To solve this problem, we propose a novel dynamic task offloading framework for distributed edge computing, leveraging the potential of meta-reinforcement learning (MRL). Our approach formulates a multi-objective optimization problem aimed at minimizing both delay and energy consumption. We model the task offloading strategy using a directed acyclic graph (DAG). Furthermore, we propose a distributed edge computing adaptive task offloading algorithm rooted in MRL. This algorithm integrates multiple Markov decision processes (MDP) with a sequence-to-sequence (seq2seq) network, enabling it to learn and adapt task offloading strategies responsively across diverse network environments. To achieve joint optimization of delay and energy consumption, we incorporate the non-dominated sorting genetic algorithm II (NSGA-II) into our framework. Simulation results demonstrate the superiority of our proposed solution, achieving a 21% reduction in time delay and a 19% decrease in energy consumption compared to alternative task offloading schemes. Moreover, our scheme exhibits remarkable adaptability, responding swiftly to changes in various network environments.
基金supported by“Regional Innovation Strategy(RIS)”through the National Research Foundation of Korea(NRF)funded by the Ministry of Education(MOE)(2023RIS-009).
文摘In a cloud environment,graphics processing units(GPUs)are the primary devices used for high-performance computation.They exploit flexible resource utilization,a key advantage of cloud environments.Multiple users share GPUs,which serve as coprocessors of central processing units(CPUs)and are activated only if tasks demand GPU computation.In a container environment,where resources can be shared among multiple users,GPU utilization can be increased by minimizing idle time because the tasks of many users run on a single GPU.However,unlike CPUs and memory,GPUs cannot logically multiplex their resources.Additionally,GPU memory does not support over-utilization:when it runs out,tasks will fail.Therefore,it is necessary to regulate the order of execution of concurrently running GPU tasks to avoid such task failures and to ensure equitable GPU sharing among users.In this paper,we propose a GPU task execution order management technique that controls GPU usage via time-based containers.The technique seeks to ensure equal GPU time among users in a container environment to prevent task failures.In the meantime,we use a deferred processing method to prevent GPU memory shortages when GPU tasks are executed simultaneously and to determine the execution order based on the GPU usage time.As the order of GPU tasks cannot be externally adjusted arbitrarily once the task commences,the GPU task is indirectly paused by pausing the container.In addition,as container pause/unpause status is based on the information about the available GPU memory capacity,overuse of GPU memory can be prevented at the source.As a result,the strategy can prevent task failure and the GPU tasks can be experimentally processed in appropriate order.
