Optimizing Data Transfers for Improved Performance on Shared GPUs Using Reinforcement Learning

Luley, Ryan; Qiu, Qinru

doi:10.1109/ccgrid.2018.00061

Cited by 2 publications

(8 citation statements)

References 7 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In comparison to such approaches, Reinforcement Learning (RL) [12] emerges as a strong alternative to solve offloading problems. However, the technical characteristics of multi-user MEC can make it challenging to apply RL, as the dimension of state and action space increases exponentially with the increasing number of users.…”

Section: A Related Workmentioning

confidence: 99%

“…in 5G scenarios), the state and action space would be very large, which poses challenges for effectively solving the optimization problem. Besides, the existing approaches require static operating models [9], [11][12] which need to be updated over time. These schemes do not scale well when the dimension of the MEC system increases with the increase the number of MEC servers and UEs associated with a larger number of configuration parameters.…”

Section: A Related Workmentioning

confidence: 99%

“…Our reputation mechanism can help overcome the limitation of the current works [11], [12], [16] in terms of mining speed. Indeed, in such works, hash power is usually allocated randomly to miners regardless of their contributions to the computation offloading process.…”

Section: A Offloading To Mec Serversmentioning

confidence: 99%

“…However, there is a possibility that this MEC server m propagates the mined block slower with other MEC servers in the propagation step, which makes such a block likely to be discarded from the blockchain. This issue is called as orphaning [12], and this miner does not receive a reward for its mining. According to [12], the orphaning probability is approximated as P orphan = 1 − e −ηφ(bm) , where η is a constant mean value and set η = 1/600(sec) [14].…”

Section: B Block Miningmentioning

confidence: 99%

“…This issue is called as orphaning [12], and this miner does not receive a reward for its mining. According to [12], the orphaning probability is approximated as P orphan = 1 − e −ηφ(bm) , where η is a constant mean value and set η = 1/600(sec) [14]. Further, b m denotes the number of transactions contained in the block that is mined by the miner m, and φ(b m ) represents the block propagation time.…”

Section: B Block Miningmentioning

confidence: 99%

See 4 more Smart Citations

Blockchain as a Service for Multi-Access Edge Computing: A Deep Reinforcement Learning Approach

Nguyen¹,

Pathirana²,

Ding³

et al. 2020

Preprint

View full text Add to dashboard Cite

Recently, blockchain has gained momentum in the academic community thanks to its decentralization, immutability, transparency and security. As an emerging paradigm, Multiaccess Edge Computing (MEC) has been widely used to provide computation and storage resources to mobile user equipments (UE) at the edge of the network for improving the performance of mobile applications. In this paper, we propose a novel blockchain-based MEC architecture where UEs can offload their computation tasks to the MEC servers. In particular, a blockchain network is deployed and hosted on the MEC platform as Blockchain as a Service (BaaS) that supports smart contract-based resource trading and transaction mining services for mobile task offloading. To enhance the performance of the blockchain-empowered MEC system, we propose a joint scheme of computation offloading and blockchain mining. Accordingly, an optimization problem is formulated to maximize edge service revenue and blockchain mining reward while minimizing the service computation latency with respect to constraints of user service demands and hash power resource. We then propose a novel Deep Reinforcement Learning (DRL) approach using a double deep Q-network (DQN) algorithm to solve the proposed problem. Numerical results demonstrate that the proposed scheme outperforms the other baseline methods in terms of better system utility with computational efficiency. Experiment results also verify that the trading contract design is efficient with low operation cost, showing the feasibility of the proposed scheme.

show abstract

Section: A Related Workmentioning

confidence: 99%