Energy-Efficient Resource Management for Federated Edge Learning With CPU-GPU Heterogeneous Computing

Abstract

Edge machine learning involves the deployment of learning algorithms at the network edge to leverage massive distributed data and computation resources to train artificial intelligence (AI) models.Among others, the framework of federated edge learning (FEEL) is popular for its data-privacy preservation. FEEL coordinates global model training at an edge server and local model training at edge devices that are connected by wireless links. This work contributes to the energy-efficient implementation of FEEL in wireless networks by designing joint computation-and-communication resource management (C 2 RM). The design targets the state-of-the-art heterogeneous mobile architecture where parallel computing using both a CPU and a GPU, called heterogeneous computing, can significantly improve both the performance and energy efficiency. To minimize the sum energy consumption of devices, we propose a novel C 2 RM framework featuring multi-dimensional control including bandwidth allocation, CPU-GPU workload partitioning and speed scaling at each device, and C 2 time division for each link. The key component of the framework is a set of equilibriums in energy rates with respect to different control variables that are proved to exist among devices or between processing units at each device. The results are applied to designing efficient algorithms for computing the optimal C 2 RM policies faster than the standard optimization tools. Based on the equilibriums, we further design energy-efficient schemes for device scheduling and greedy spectrum sharing that scavenges "spectrum holes" resulting from heterogeneous C 2 time divisions among devices. Using a real dataset, experiments are conducted to demonstrate the effectiveness of C 2 RM on improving the energy efficiency of a FEEL system.

EOP: efficient operator partition for deep learning inference over edge servers

“…In DL inference scenarios, each operator before partitioning needs an exclusive resource (supported by TVM), and all operators run based on common default settings of mainstream DL frameworks [15,23,27,29,30]. After partitioning, some operators can be partitioned into two sub-operators on edge servers that can improve DL inference performance.…”

“…EOP improves DL inference performance on edge servers by: 1) accurately estimating operator execution on heterogeneous resources, and 2) efficiently partitioning key operators. Note that a lot of work are mainly on edge-cloud collaborations [11,29,30] and EOP is non-intrusion with them. As such, we discuss the related work from these two research aspects.…”

“…Running various Deep Learning (DL) models, such as Convolutional Neural Network (CNN) [22], Recurrent Neural Network (RNN) [20], Yolo [21] and Transformers [25] over edge servers can empower artificial intelligence Internet of Things applications. It has been proofed as an important way to achieve significant performance improvement of DL models [15,23,30]. In this paper, we focus on partition optimization to run partitioned operators in parallel to use heterogeneous resource within an edge server.…”

“…They still suffer from a set of limitations: the assumption of operator partition on CPU and GPU with comparable computational throughput [15], or only on homogeneous devices [27] results in equal partition ratios (each CPU and GPU have 1/2 tensors) and may make CPU/GPU idle. Manually setting fixed operator partition ratios for different operators or subgraphs [23,29,30] can only work for heterogeneous resources with same computation architecture (e.g., TPUv2/v3), which cannot be used in CPU/GPU heterogeneous computation (LLVM in CPU and CUDA in GPU).…”

Editorial Recent Techniques of Green Information and Communications Technologies

“…While current federated ML algorithms mostly aim at improving the prediction accuracy of obtained models, a few researchers have considered the carbon footprint of the training process. For example, Zeng et al [13] investigated the energy consumption minimization problem by joint management of the computation and communication resources. Their algorithm aims to achieve an equilibrium among the four dimensions of resources, i.e., bandwidth, CPU-GPU operating period, CPU-GPU frequency, and CPU-GPU workload partition.…”