Edge-Facilitated Wireless Distributed Computing
Edge-Facilitated Wireless Distributed Computing
复制标题
边缘促进的无线分布式计算
DOI：
--
复制
发表时间：
2016
期刊：
影响因子：
--
通讯作者：
A. Avestimehr
中科院分区：
文献类型：
--
作者：
Songze Li;Qian Yu;M. Maddah;A. Avestimehr
我们提出了一个框架，边缘便利的无线分布式计算，其中几个移动的用户连接到一个接入点协作的分布式计算任务。我们的特点是最小的通信负载，无论是在上行链路（从用户到接入点）和下行链路（从接入点到用户），需要分布式计算。特别是，我们开发了一种通信方案和数据集放置策略，该策略在用户处引起特定的计算重叠，然后可以利用该重叠在用户和接入点处进行编码，以显着降低通信负载。我们证明了通信负载的减少（与未编码的解决方案相比）可以随着网络的大小线性扩展（即，用户的数量），因此我们提出的方案可以导致用于边缘促进的无线分布式计算的“可缩放”设计（即，容纳任意数量的用户而不引起额外的通信负载）。此外，我们通过建立严格的信息论外界来确定所提出的方案的最优性，并证明所提出的方案同时实现了最小的上行链路和下行链路通信负载。我们还将结果推广到一个分散的设置，其中一个随机的和先验未知的用户子集可能会参与分布式计算在每个时间，并表征均匀随机数据集放置在用户的最小通信负载。
We propose a framework for edge-facilitated wireless distributed computing, in which several mobile users connected to an access point collaborate for a distributed computing task. We characterize the minimum communication load, both in uplink (from users to the access point) and downlink (from access point to the users), required for distributed computing. In particular, we develop a communication scheme and a dataset placement strategy that induces a particular overlap of computations at the users, which can then be exploited for coding at both users and the access point to significantly reduce the communication load. We demonstrate that the reduction in communication load (compared to uncoded solutions) can scale linearly with the size of the network (i.e., the number of users), hence our proposed scheme can result in a "scalable" design for edge- facilitated wireless distributed computing (i.e., accommodating any number of users without incurring extra communication load). Furthermore, we establish the optimality of the proposed scheme by developing a tight information theoretic outer- bound, and demonstrate that the proposed scheme achieves the minimum uplink and downlink communication load simultaneously. We also generalize the results to a decentralized setting, in which a random and a priori unknown subset of users may participate in distributed computing at each time, and characterize the minimum communication load for uniformly random dataset placement at users.