Training deep neural networks (DNNs) is a major workload in datacenters today, resulting in a tremendously fast growth of energy consumption. It is important to reduce the energy consumption while completing the DL training jobs early in data centers. In this paper, we propose PowerFlow, a GPU clusters scheduler that reduces the average Job Completion Time (JCT) under an energy budget. We first present performance models for DL training jobs to predict the throughput and energy consumption performance with different configurations. Based on the performance models, PowerFlow dynamically allocates GPUs and adjusts the GPU-level or job-level configurations of DL training jobs. PowerFlow applies network packing and buddy allocation to job placement, thus avoiding extra energy consumed by cluster fragmentations. Evaluation results show that under the same energy consumption, PowerFlow improves the average JCT by 1.57 - 3.39 x at most, compared to competitive baselines.
相關內容
This paper addresses the scheduling problem on two identical parallel machines with a single server in charge of loading and unloading operations of jobs. Each job has to be loaded by the server before being processed on one of the two machines and unloaded by the same server after its processing. No delay is allowed between loading and processing, and between processing and unloading. The objective function involves the minimization of the makespan. This problem referred to as P2, S1|sj , tj |Cmax generalizes the classical parallel machine scheduling problem with a single server which performs only the loading (i.e., setup) operation of each job. For this NP-hard problem, no solution algorithm was proposed in the literature. Therefore, we present two mixedinteger linear programming (MILP) formulations, one with completion-time variables along with two valid inequalities and one with time-indexed variables. In addition, we propose some polynomial-time solvable cases and a tight theoretical lower bound. In addition, we show that the minimization of the makespan is equivalent to the minimization of the total idle times on the machines. To solve large-sized instances of the problem, an efficient General Variable Neighborhood Search (GVNS) metaheuristic with two mechanisms for finding an initial solution is designed. The GVNS is evaluated by comparing its performance with the results provided by the MILPs and another metaheuristic. The results show that the average percentage deviation from the theoretical lower-bound of GVNS is within 0.642%. Some managerial insights are presented and our results are compared with the related literature.
The primary operation in DNNs is the dot product of quantized input activations and weights. Prior works have proposed the design of memory-centric architectures based on the Processing-In-Memory (PIM) paradigm. Resistive RAM (ReRAM) technology is especially appealing for PIM-based DNN accelerators due to its high density to store weights, low leakage energy, low read latency, and high performance capabilities to perform the DNN dot-products massively in parallel within the ReRAM crossbars. However, the main bottleneck of these architectures is the energy-hungry analog-to-digital conversions (ADCs) required to perform analog computations in-ReRAM, which penalizes the efficiency and performance benefits of PIM. To improve energy-efficiency of in-ReRAM analog dot-product computations we present ReDy, a hardware accelerator that implements a ReRAM-centric Dynamic quantization scheme to take advantage of the bit serial streaming and processing of activations. The energy consumption of ReRAM-based DNN accelerators is directly proportional to the numerical precision of the input activations of each DNN layer. In particular, ReDy exploits that activations of CONV layers from Convolutional Neural Networks (CNNs), a subset of DNNs, are commonly grouped according to the size of their filters and the size of the ReRAM crossbars. Then, ReDy quantizes on-the-fly each group of activations with a different numerical precision based on a novel heuristic that takes into account the statistical distribution of each group. Overall, ReDy greatly reduces the activity of the ReRAM crossbars and the number of A/D conversions compared to an static 8-bit uniform quantization. We evaluate ReDy on a popular set of modern CNNs. On average, ReDy provides 13\% energy savings over an ISAAC-like accelerator with negligible accuracy loss and area overhead.
The current trend in developing machine learning models for reading comprehension and logical reasoning tasks is focused on improving the models' abilities to understand and utilize logical rules. This work focuses on providing a novel loss function and accompanying model architecture that has more interpretable components than some other models by representing a common strategy employed by humans when given reading comprehension and logical reasoning tasks. This strategy involves emphasizing relative accuracy over absolute accuracy and can theoretically produce the correct answer without full knowledge of the information required to solve the question. We examine the effectiveness of applying such a strategy to train transfer learning models to solve reading comprehension and logical reasoning questions. The models were evaluated on the ReClor dataset, a challenging reading comprehension and logical reasoning benchmark. We propose the polytuplet loss function, an extension of the triplet loss function, to ensure prioritization of learning the relative correctness of answer choices over learning the true accuracy of each choice. Our results indicate that models employing polytuplet loss outperform existing baseline models. Although polytuplet loss is a promising alternative to other contrastive loss functions, further research is required to quantify the benefits it may present.
Spiking Neural Networks (SNNs) are biologically plausible models that have been identified as potentially apt for the deployment for energy-efficient intelligence at the edge, particularly for sequential learning tasks. However, training of SNNs poses a significant challenge due to the necessity for precise temporal and spatial credit assignment. Back-propagation through time (BPTT) algorithm, whilst being the most widely used method for addressing these issues, incurs a high computational cost due to its temporal dependency. Moreover, BPTT and its approximations solely utilize causal information derived from the spiking activity to compute the synaptic updates, thus neglecting non-causal relationships. In this work, we propose S-TLLR, a novel three-factor temporal local learning rule inspired by the Spike-Timing Dependent Plasticity (STDP) mechanism, aimed at training SNNs on event-based learning tasks. S-TLLR considers both causal and non-causal relationships between pre and post-synaptic activities, achieving performance comparable to BPTT and enhancing performance relative to methods using only causal information. Furthermore, S-TLLR has low memory and time complexity, which is independent of the number of time steps, rendering it suitable for online learning on low-power devices. To demonstrate the scalability of our proposed method, we have conducted extensive evaluations on event-based datasets spanning a wide range of applications, such as image and gesture recognition, audio classification, and optical flow estimation. In all the experiments, S-TLLR achieved high accuracy with a reduction in the number of computations between $1.1-10\times$.
This paper presents new methods for analyzing and evaluating generalized plans that can solve broad classes of related planning problems. Although synthesis and learning of generalized plans has been a longstanding goal in AI, it remains challenging due to fundamental gaps in methods for analyzing the scope and utility of a given generalized plan. This paper addresses these gaps by developing a new conceptual framework along with proof techniques and algorithmic processes for assessing termination and goal-reachability related properties of generalized plans. We build upon classic results from graph theory to decompose generalized plans into smaller components that are then used to derive hierarchical termination arguments. These methods can be used to determine the utility of a given generalized plan, as well as to guide the synthesis and learning processes for generalized plans. We present theoretical as well as empirical results illustrating the scope of this new approach. Our analysis shows that this approach significantly extends the class of generalized plans that can be assessed automatically, thereby reducing barriers in the synthesis and learning of reliable generalized plans.
The adaptive processing of structured data is a long-standing research topic in machine learning that investigates how to automatically learn a mapping from a structured input to outputs of various nature. Recently, there has been an increasing interest in the adaptive processing of graphs, which led to the development of different neural network-based methodologies. In this thesis, we take a different route and develop a Bayesian Deep Learning framework for graph learning. The dissertation begins with a review of the principles over which most of the methods in the field are built, followed by a study on graph classification reproducibility issues. We then proceed to bridge the basic ideas of deep learning for graphs with the Bayesian world, by building our deep architectures in an incremental fashion. This framework allows us to consider graphs with discrete and continuous edge features, producing unsupervised embeddings rich enough to reach the state of the art on several classification tasks. Our approach is also amenable to a Bayesian nonparametric extension that automatizes the choice of almost all model's hyper-parameters. Two real-world applications demonstrate the efficacy of deep learning for graphs. The first concerns the prediction of information-theoretic quantities for molecular simulations with supervised neural models. After that, we exploit our Bayesian models to solve a malware-classification task while being robust to intra-procedural code obfuscation techniques. We conclude the dissertation with an attempt to blend the best of the neural and Bayesian worlds together. The resulting hybrid model is able to predict multimodal distributions conditioned on input graphs, with the consequent ability to model stochasticity and uncertainty better than most works. Overall, we aim to provide a Bayesian perspective into the articulated research field of deep learning for graphs.
As soon as abstract mathematical computations were adapted to computation on digital computers, the problem of efficient representation, manipulation, and communication of the numerical values in those computations arose. Strongly related to the problem of numerical representation is the problem of quantization: in what manner should a set of continuous real-valued numbers be distributed over a fixed discrete set of numbers to minimize the number of bits required and also to maximize the accuracy of the attendant computations? This perennial problem of quantization is particularly relevant whenever memory and/or computational resources are severely restricted, and it has come to the forefront in recent years due to the remarkable performance of Neural Network models in computer vision, natural language processing, and related areas. Moving from floating-point representations to low-precision fixed integer values represented in four bits or less holds the potential to reduce the memory footprint and latency by a factor of 16x; and, in fact, reductions of 4x to 8x are often realized in practice in these applications. Thus, it is not surprising that quantization has emerged recently as an important and very active sub-area of research in the efficient implementation of computations associated with Neural Networks. In this article, we survey approaches to the problem of quantizing the numerical values in deep Neural Network computations, covering the advantages/disadvantages of current methods. With this survey and its organization, we hope to have presented a useful snapshot of the current research in quantization for Neural Networks and to have given an intelligent organization to ease the evaluation of future research in this area.
Spectral clustering (SC) is a popular clustering technique to find strongly connected communities on a graph. SC can be used in Graph Neural Networks (GNNs) to implement pooling operations that aggregate nodes belonging to the same cluster. However, the eigendecomposition of the Laplacian is expensive and, since clustering results are graph-specific, pooling methods based on SC must perform a new optimization for each new sample. In this paper, we propose a graph clustering approach that addresses these limitations of SC. We formulate a continuous relaxation of the normalized minCUT problem and train a GNN to compute cluster assignments that minimize this objective. Our GNN-based implementation is differentiable, does not require to compute the spectral decomposition, and learns a clustering function that can be quickly evaluated on out-of-sample graphs. From the proposed clustering method, we design a graph pooling operator that overcomes some important limitations of state-of-the-art graph pooling techniques and achieves the best performance in several supervised and unsupervised tasks.
Deep neural networks (DNNs) are successful in many computer vision tasks. However, the most accurate DNNs require millions of parameters and operations, making them energy, computation and memory intensive. This impedes the deployment of large DNNs in low-power devices with limited compute resources. Recent research improves DNN models by reducing the memory requirement, energy consumption, and number of operations without significantly decreasing the accuracy. This paper surveys the progress of low-power deep learning and computer vision, specifically in regards to inference, and discusses the methods for compacting and accelerating DNN models. The techniques can be divided into four major categories: (1) parameter quantization and pruning, (2) compressed convolutional filters and matrix factorization, (3) network architecture search, and (4) knowledge distillation. We analyze the accuracy, advantages, disadvantages, and potential solutions to the problems with the techniques in each category. We also discuss new evaluation metrics as a guideline for future research.
Spectral clustering is a leading and popular technique in unsupervised data analysis. Two of its major limitations are scalability and generalization of the spectral embedding (i.e., out-of-sample-extension). In this paper we introduce a deep learning approach to spectral clustering that overcomes the above shortcomings. Our network, which we call SpectralNet, learns a map that embeds input data points into the eigenspace of their associated graph Laplacian matrix and subsequently clusters them. We train SpectralNet using a procedure that involves constrained stochastic optimization. Stochastic optimization allows it to scale to large datasets, while the constraints, which are implemented using a special-purpose output layer, allow us to keep the network output orthogonal. Moreover, the map learned by SpectralNet naturally generalizes the spectral embedding to unseen data points. To further improve the quality of the clustering, we replace the standard pairwise Gaussian affinities with affinities leaned from unlabeled data using a Siamese network. Additional improvement can be achieved by applying the network to code representations produced, e.g., by standard autoencoders. Our end-to-end learning procedure is fully unsupervised. In addition, we apply VC dimension theory to derive a lower bound on the size of SpectralNet. State-of-the-art clustering results are reported on the Reuters dataset. Our implementation is publicly available at //github.com/kstant0725/SpectralNet .
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191