Collaborative inference has received significant research interest in machine learning as a vehicle for distributing computation load, reducing latency, as well as addressing privacy preservation in communications. Recent collaborative inference frameworks have adopted dynamic inference methodologies such as early-exit and run-time partitioning of neural networks. However, as machine learning frameworks scale in the number of inference inputs, e.g., in surveillance applications, fault tolerance related to device failure needs to be considered. This paper presents the Edge-PRUNE distributed computing framework, built on a formally defined model of computation, which provides a flexible infrastructure for fault tolerant collaborative inference. The experimental section of this work shows results on achievable inference time savings by collaborative inference, presents fault tolerant system topologies and analyzes their cost in terms of execution time overhead.
We develop an eigenspace estimation algorithm for distributed environments with arbitrary node failures, where a subset of computing nodes can return structurally valid but otherwise arbitrarily chosen responses. Notably, this setting encompasses several important scenarios that arise in distributed computing and data-collection environments such as silent/soft errors, outliers or corrupted data at certain nodes, and adversarial responses. Our estimator builds upon and matches the performance of a recently proposed non-robust estimator up to an additive $\tilde{O}(\sigma \sqrt{\alpha})$ error, where $\sigma^2$ is the variance of the existing estimator and $\alpha$ is the fraction of corrupted nodes.
This paper studies asynchronous message passing (AMP), a new paradigm for applying neural network based learning to graphs. Existing graph neural networks use the synchronous distributed computing model and aggregate their neighbors in each round, which causes problems such as oversmoothing and limits their expressiveness. On the other hand, AMP is based on the asynchronous model, where nodes react to messages of their neighbors individually. We prove that (i) AMP can simulate synchronous GNNs and that (ii) AMP can theoretically distinguish any pair of graphs. We experimentally validate AMP's expressiveness. Further, we show that AMP might be better suited to propagate messages over large distances in graphs and performs well on several graph classification benchmarks.
Distributed computing, such as cloud computing, provides promising platforms to execute multiple workflows. Workflow scheduling plays an important role in multi-workflow execution with multi-objective requirements. Although there exist many multi-objective scheduling algorithms, they focus mainly on optimizing makespan and cost for a single workflow. There is a limited research on multi-objective optimization for multi-workflow scheduling. Considering multi-workflow scheduling, there is an additional key objective to maintain the fairness of workflows using the resources. To address such issues, this paper first defines a new multi-objective optimization model based on makespan, cost, and fairness, and then proposes a global clustering-based multi-workflow scheduling strategy for resource allocation. Experimental results show that the proposed approach performs better than the compared algorithms without significant compromise of the overall makespan and cost as well as individual fairness, which can guide the simulation workflow scheduling on clouds.
With massive penetrations of active grid-edge technologies, distributed computing and optimization paradigm has gained significant attention to solve distribution-level optimal power flow (OPF) problems. However, the application of generic distributed optimization techniques to OPF problems leads to a very large number of macro-iterations or communication rounds among the distributed computing agents delaying the decision-making process or resulting in suboptimal solutions. Moreover, the existing distribution-level OPF problems typically model inverter-interfaced distributed energy resources (DERs) as grid-following inverters; grid-supporting and grid-forming functionalities have not been explicitly considered. The added complexities introduced by different inverter models require further attention to developing an appropriate model for new types of inverter-based DERs and computationally-tractable OPF algorithms. In this paper, we expand the distribution-level OPF model to include a combination of the grid-forming, grid-supporting, grid-following inverter-based DERs and also present the application of a domain-specific problem decomposition and distributed algorithm for the topologically radial power distribution systems to efficiently solve distribution-level OPF problem.
While modern internet services, such as chatbots, search engines, and online advertising, demand the use of large-scale deep neural networks (DNNs), distributed training and inference over heterogeneous computing systems are desired to facilitate these DNN models. Mixture-of-Experts (MoE) is one the most common strategies to lower the cost of training subject to the overall size of models/data through gating and parallelism in a divide-and-conquer fashion. While DeepSpeed has made efforts in carrying out large-scale MoE training over heterogeneous infrastructures, the efficiency of training and inference could be further improved from several system aspects, including load balancing, communication/computation efficiency, and memory footprint limits. In this work, we present a novel MoESys that boosts efficiency in both large-scale training and inference. Specifically, in the training procedure, the proposed MoESys adopts an Elastic MoE training strategy with 2D prefetch and Fusion communication over Hierarchical storage, so as to enjoy efficient parallelisms. For scalable inference in a single node, especially when the model size is larger than GPU memory, MoESys builds the CPU-GPU memory jointly into a ring of sections to load the model, and executes the computation tasks across the memory sections in a round-robin manner for efficient inference. We carried out extensive experiments to evaluate MoESys, where MoESys successfully trains a Unified Feature Optimization (UFO) model with a Sparsely-Gated Mixture-of-Experts model of 12B parameters in 8 days on 48 A100 GPU cards. The comparison against the state-of-the-art shows that MoESys outperformed DeepSpeed with 33% higher throughput (tokens per second) in training and 13% higher throughput in inference in general. Particularly, under unbalanced MoE Tasks, e.g., UFO, MoESys achieved 64% higher throughput with 18% lower memory footprints.
The ever-growing model size and scale of compute have attracted increasing interests in training deep learning models over multiple nodes. However, when it comes to training on cloud clusters, especially across remote clusters, huge challenges are faced. In this work, we introduce a general framework, Nebula-I, for collaboratively training deep learning models over remote heterogeneous clusters, the connections between which are low-bandwidth wide area networks (WANs). We took natural language processing (NLP) as an example to show how Nebula-I works in different training phases that include: a) pre-training a multilingual language model using two remote clusters; and b) fine-tuning a machine translation model using knowledge distilled from pre-trained models, which run through the most popular paradigm of recent deep learning. To balance the accuracy and communication efficiency, in Nebula-I, parameter-efficient training strategies, hybrid parallel computing methods and adaptive communication acceleration techniques are jointly applied. Meanwhile, security strategies are employed to guarantee the safety, reliability and privacy in intra-cluster computation and inter-cluster communication. Nebula-I is implemented with the PaddlePaddle deep learning framework, which can support collaborative training over heterogeneous hardware, e.g. GPU and NPU. Experiments demonstrate that the proposed framework could substantially maximize the training efficiency while preserving satisfactory NLP performance. By using Nebula-I, users can run large-scale training tasks over cloud clusters with minimum developments, and the utility of existed large pre-trained models could be further promoted. We also introduced new state-of-the-art results on cross-lingual natural language inference tasks, which are generated based upon a novel learning framework and Nebula-I.
In this work, we present a distributed framework based on the graph algorithm for computing control invariant set for nonlinear cascade systems. The proposed algorithm exploits the structure of the interconnections within a process network. First, the overall system is decomposed into several subsystems with overlapping states. Second, the control invariant set for the subsystems are computed in a distributed manner. Finally, an approximation of the control invariant set for the overall system is reconstructed from the subsystem solutions and validated. We demonstrate the efficacy and convergence of the proposed method to the centralized graph-based algorithm using several numerical examples including a six dimensional continuous stirred tank reactor system.
The Metaverse is a virtual environment where users are represented by avatars to navigate a virtual world, which has strong links with the physical one. State-of-the-art Metaverse architectures rely on a cloud-based approach for avatar physics emulation and graphics rendering computation. Such centralized design is unfavorable as it suffers from several drawbacks caused by the long latency required for cloud access, such as low quality visualization. To solve this issue, in this paper, we propose a Fog-Edge hybrid computing architecture for Metaverse applications that leverage an edge-enabled distributed computing paradigm, which makes use of edge devices computing power to fulfil the required computational cost for heavy tasks such as collision detection in virtual universe and computation of 3D physics in virtual simulation. The computational cost related to an entity in the Metaverse such as collision detection or physics emulation are performed at the end-device of the associated physical entity. To prove the effectiveness of the proposed architecture, we simulate a distributed social metaverse application. Simulation results shows that the proposed architecture can reduce the latency by 50% when compared with the legacy cloud-based Metaverse applications.
When dealing with the Internet of Things (IoT), especially industrial IoT (IIoT), two manifest challenges leap to mind. First is the massive amount of data streaming to and from IoT devices, and second is the fast pace at which these systems must operate. Distributed computing in the form of edge/cloud structure is a popular technique to overcome these two challenges. In this paper, we propose ADDAI (Anomaly Detection using Distributed AI) that can easily span out geographically to cover a large number of IoT sources. Due to its distributed nature, it guarantees critical IIoT requirements such as high speed, robustness against a single point of failure, low communication overhead, privacy, and scalability. Through empirical proof, we show the communication cost is minimized, and the performance improves significantly while maintaining the privacy of raw data at the local layer. ADDAI provides predictions for new random samples with an average success rate of 98.4% while reducing the communication overhead by half compared with the traditional technique of offloading all the raw sensor data to the cloud.
We consider the problem of evaluating arbitrary multivariate polynomials over a massive dataset containing multiple inputs, on a distributed computing system with a master node and multiple worker nodes. Generalized Lagrange Coded Computing (GLCC) codes are proposed to simultaneously provide resiliency against stragglers who do not return computation results in time, security against adversarial workers who deliberately modify results for their benefit, and information-theoretic privacy of the dataset amidst possible collusion of workers. GLCC codes are constructed by first partitioning the dataset into multiple groups, then encoding the dataset using carefully designed interpolating polynomials, and sharing multiple encoded data points to each worker, such that interference computation results across groups can be eliminated at the master. Particularly, GLCC codes include the state-of-the-art Lagrange Coded Computing (LCC) codes as a special case, and exhibit a more flexible tradeoff between communication and computation overheads in optimizing system efficiency. Furthermore, we apply GLCC to distributed training of machine learning models, and demonstrate that GLCC codes achieve a speedup of up to $2.5\text{--}3.9\times$ over LCC codes in training time, across experiments for training image classifiers on different datasets, model architectures, and straggler patterns.
We design deep neural networks (DNNs) and corresponding networks' splittings to distribute DNNs' workload to camera sensors and a centralized aggregator on head mounted devices to meet system performance targets in inference accuracy and latency under the given hardware resource constraints. To achieve an optimal balance among computation, communication, and performance, a split-aware neural architecture search framework, SplitNets, is introduced to conduct model designing, splitting, and communication reduction simultaneously. We further extend the framework to multi-view systems for learning to fuse inputs from multiple camera sensors with optimal performance and systemic efficiency. We validate SplitNets for single-view system on ImageNet as well as multi-view system on 3D classification, and show that the SplitNets framework achieves state-of-the-art (SOTA) performance and system latency compared with existing approaches.
Fashion is now among the largest industries worldwide, for it represents human history and helps tell the worlds story. As a result of the Fourth Industrial Revolution, the Internet has become an increasingly important source of fashion information. However, with a growing number of web pages and social data, it is nearly impossible for humans to manually catch up with the ongoing evolution and the continuously variable content in this domain. The proper management and exploitation of big data can pave the way for the substantial growth of the global economy as well as citizen satisfaction. Therefore, computer scientists have found it challenging to handle e-commerce fashion websites by using big data and machine learning technologies. This paper first proposes a scalable focused Web Crawler engine based on the distributed computing platforms to extract and process fashion data on e-commerce websites. The role of the proposed platform is then described in developing a disentangled feature extraction method by employing deep convolutional generative adversarial networks (DCGANs) for content-based image indexing and retrieval. Finally, the state-of-the-art solutions are compared, and the results of the proposed approach are analyzed on a standard dataset. For the real-life implementation of the proposed solution, a Web-based application is developed on Apache Storm, Kafka, Solr, and Milvus platforms to create a fashion search engine called SnapMode.
Most AI projects start with a Python notebook running on a single laptop; however, one usually needs to go through a mountain of pains to scale it to handle larger dataset (for both experimentation and production deployment). These usually entail many manual and error-prone steps for the data scientists to fully take advantage of the available hardware resources (e.g., SIMD instructions, multi-processing, quantization, memory allocation optimization, data partitioning, distributed computing, etc.). To address this challenge, we have open sourced BigDL 2.0 at https://github.com/intel-analytics/BigDL/ under Apache 2.0 license (combining the original BigDL and Analytics Zoo projects); using BigDL 2.0, users can simply build conventional Python notebooks on their laptops (with possible AutoML support), which can then be transparently accelerated on a single node (with up-to 9.6x speedup in our experiments), and seamlessly scaled out to a large cluster (across several hundreds servers in real-world use cases). BigDL 2.0 has already been adopted by many real-world users (such as Mastercard, Burger King, Inspur, etc.) in production.
Federated learning (FL) is a key enabler for efficient communication and computing, leveraging devices' distributed computing capabilities. However, applying FL in practice is challenging due to the local devices' heterogeneous energy, wireless channel conditions, and non-independently and identically distributed (non-IID) data distributions. To cope with these issues, this paper proposes a novel learning framework by integrating FL and width-adjustable slimmable neural networks (SNN). Integrating FL with SNNs is challenging due to time-varying channel conditions and data distributions. In addition, existing multi-width SNN training algorithms are sensitive to the data distributions across devices, which makes SNN ill-suited for FL. Motivated by this, we propose a communication and energy-efficient SNN-based FL (named SlimFL) that jointly utilizes superposition coding (SC) for global model aggregation and superposition training (ST) for updating local models. By applying SC, SlimFL exchanges the superposition of multiple-width configurations decoded as many times as possible for a given communication throughput. Leveraging ST, SlimFL aligns the forward propagation of different width configurations while avoiding inter-width interference during backpropagation. We formally prove the convergence of SlimFL. The result reveals that SlimFL is not only communication-efficient but also deals with non-IID data distributions and poor channel conditions, which is also corroborated by data-intensive simulations.
For intelligent home IoT services with sensors and machine learning, we need to upload IoT data to the cloud server which cannot share private data for training. A recent machine learning approach, called federated learning, keeps user data on the device in the distributed computing environment. Though federated learning is useful for protecting privacy, it experiences poor performance in terms of the end-to-end response time in home IoT services, because IoT devices are usually controlled by remote servers in the cloud. In addition, it is difficult to achieve the high accuracy of federated learning models due to insufficient data problems and model inversion attacks. In this paper, we propose a local IoT control method for a federated learning home service that recognizes the user behavior in the home network quickly and accurately. We present a federated learning client with transfer learning and differential privacy to solve data scarcity and data model inversion attack problems. From experiments, we show that the local control of home IoT devices for user authentication and control message transmission by the federated learning clients improves the response time to less than 1 second. Moreover, we demonstrate that federated learning with transfer learning achieves 97% of accuracy under 9,000 samples, which is only 2% of the difference from centralized learning.
Federated learning is proposed by Google to safeguard data privacy through training models locally on users' devices. However, with deep learning models growing in size to achieve better results, it becomes increasingly difficult to accommodate the whole model on one single device. Thus, model parallelism is then used to divide the model weights among several devices. With this logic, the approach currently used evenly allocates weights among devices. However, in reality, a computation bottleneck may occur resulting from variant computing power of different users' devices. To address this problem, load balancing is needed to allocate the model weights based on the computational capability of the device. In this paper, we proposed Sky Computing, a load-balanced model parallelism framework to adaptively allocate the weights to devices. Sky Computing outperforms the baseline method by 55% in training time when training 160-layer BERT with 64 nodes. The source code can be found at https://github.com/hpcaitech/SkyComputing.
Most multi-agent reinforcement learning (MARL) methods are limited in the scale of problems they can handle. With increasing numbers of agents, the number of training iterations required to find the optimal behaviors increases exponentially due to the exponentially growing joint state and action spaces. This paper tackles this limitation by introducing a scalable MARL method called Distributed multi-Agent Reinforcement Learning with One-hop Neighbors (DARL1N). DARL1N is an off-policy actor-critic method that addresses the curse of dimensionality by restricting information exchanges among the agents to one-hop neighbors when representing value and policy functions. Each agent optimizes its value and policy functions over a one-hop neighborhood, significantly reducing the learning complexity, yet maintaining expressiveness by training with varying neighbor numbers and states. This structure allows us to formulate a distributed learning framework to further speed up the training procedure. Distributed computing systems, however, contain straggler compute nodes, which are slow or unresponsive due to communication bottlenecks, software or hardware problems. To mitigate the detrimental straggler effect, we introduce a novel coded distributed learning architecture, which leverages coding theory to improve the resilience of the learning system to stragglers. Comprehensive experiments show that DARL1N significantly reduces training time without sacrificing policy quality and is scalable as the number of agents increases. Moreover, the coded distributed learning architecture improves training efficiency in the presence of stragglers.
As easy-to-use deep learning libraries such as Tensorflow and Pytorch are popular, it has become convenient to develop machine learning models. Due to privacy issues with centralized machine learning, recently, federated learning in the distributed computing framework is attracting attention. The central server does not collect sensitive and personal data from clients in federated learning, but it only aggregates the model parameters. Though federated learning helps protect privacy, it is difficult for machine learning developers to share the models that they could utilize for different-domain applications. In this paper, we propose a federated learning model sharing service named Federated Learning Hub (FLHub). Users can upload, download, and contribute the model developed by other developers similarly to GitHub. We demonstrate that a forked model can finish training faster than the existing model and that learning progressed more quickly for each federated round.
Tensors are a fundamental operation in distributed computing, \emph{e.g.,} machine learning, that are commonly distributed into multiple parallel tasks for large datasets. Stragglers and other failures can severely impact the overall completion time. Recent works in coded computing provide a novel strategy to mitigate stragglers with coded tasks, with an objective of minimizing the number of tasks needed to recover the overall result, known as the recovery threshold. However, we demonstrate that this strict combinatorial definition does not directly optimize the probability of failure. In this paper, we focus on the most likely event and measure the optimality of a coding scheme more directly by its probability of decoding. Our probabilistic approach leads us to a practical construction of random codes for matrix multiplication, i.e., locally random alloy codes, which are optimal with respect to the measures. Furthermore, the probabilistic approach allows us to discover a surprising impossibility theorem about both random and deterministic coded distributed tensors.