Reduce Memory Usage Research Articles

CASS: A distributed network clustering algorithm based on structure similarity for large-scale network.

As the size of networks increases, it is becoming important to analyze large-scale network data. A network clustering algorithm is useful for analysis of network data. Conventional network clustering algorithms in a single machine environment rather than a parallel machine environment are actively being researched. However, these algorithms cannot analyze large-scale network data because of memory size issues. As a solution, we propose a network clustering algorithm for large-scale network data analysis using Apache Spark by changing the paradigm of the conventional clustering algorithm to improve its efficiency in the Apache Spark environment. We also apply optimization approaches such as Bloom filter and shuffle selection to reduce memory usage and execution time. By evaluating our proposed algorithm based on an average normalized cut, we confirmed that the algorithm can analyze diverse large-scale network datasets such as biological, co-authorship, internet topology and social networks. Experimental results show that the proposed algorithm can develop more accurate clusters than comparative algorithms with less memory usage. Furthermore, we confirm the proposed optimization approaches and the scalability of the proposed algorithm. In addition, we validate that clusters found from the proposed algorithm can represent biologically meaningful functions.

Efficient diffraction calculations using implicit convolution

Diffraction calculations, such as Fresnel diffraction and the angular spectrum method, are essential in optical simulations. These methods are numerically calculated by discrete Fourier transform or fast Fourier transform (FFT). Although these diffraction calculations require FFT-based convolution, FFT-based convolution becomes circular convolution without using zero padding. Zero padding helps to avoid circular convolution, but increases calculation time and memory usage. This paper proposes efficient diffraction calculations using implicit convolution. The proposed diffraction scheme accelerates calculation time, reduces memory usage, and can be applied to any convolution-based diffraction.

A Novel ADE-CN-FDTD With Improved Computational Efficiency for Dispersive Media

A novel auxiliary difference equation (ADE) finite-difference time-domain (FDTD) method with Crank–Nicolson (CN) scheme is proposed for electromagnetic simulation of dispersive materials in this letter. The proposed method introduces an ADE technique that establishes the relationship between the electric displacement vector and electric field intensity in dispersive media with a difference equation. Hence, the CN scheme is only applied to one of the Maxwell curl equations, resulting in reduced memory usage and computing time. To validate the advantages of the proposed approach, an example with plane wave propagation in a dispersive media is presented. Compared with the conventional ADE-CN-FDTD method, the proposed method consumes much less computing time while preserves excellent simulation accuracy.

Efficient Data-Driven Geologic Feature Characterization from Pre-stack Seismic Measurements using Randomized Machine-Learning Algorithm

Conventional seismic techniques for detecting the subsurface geologic features are challenged by limited data coverage, computational inefficiency, and subjective human factors. We developed a novel data-driven geological feature detection approach based on pre-stack seismic measurements. Our detection method employs an efficient and accurate machine-learning detection approach to extract useful subsurface geologic features automatically. Specifically, our method is based on kernel ridge regression model. The conventional kernel ridge regression can be computationally prohibited because of the large volume of seismic measurements. We employ a data reduction technique in combination with the conventional kernel ridge regression method to improve the computational efficiency and reduce memory usage. In particular, we utilize a randomized numerical linear algebra technique, named Nystr\"om method, to effectively reduce the dimensionality of the feature space without compromising the information content required for accurate detection. We provide thorough computational cost analysis to show efficiency of our new geological feature detection methods. We further validate the performance of our new subsurface geologic feature detection method using synthetic surface seismic data for 2D acoustic and elastic velocity models. Our numerical examples demonstrate that our new detection method significantly improves the computational efficiency while maintaining comparable accuracy. Interestingly, we show that our method yields a speed-up ratio on the order of $\sim10^2$ to $\sim 10^3$ in a multi-core computational environment.

Strict Very Fast Decision Tree: A memory conservative algorithm for data stream mining

Dealing with memory and time constraints are current challenges when learning from data streams with a massive amount of data. Many algorithms have been proposed to handle these difficulties, among them, the Very Fast Decision Tree (VFDT) algorithm. Although the VFDT has been widely used in data stream mining, in the last years, several authors have suggested modifications to increase its performance, putting aside memory concerns by proposing memory-costly solutions. Besides, most data stream mining solutions have been centred around ensembles, which combine the memory costs of their weak learners, usually VFDTs. To reduce memory costs, keeping predictive performance, this study proposes the Strict VFDT (SVFDT), a novel algorithm based on the VFDT. The SVFDT algorithm minimises unnecessary tree growth, substantially reducing memory usage and keeping competitive predictive performance. Moreover, since it creates much more shallow trees than VFDT, the SVFDT can achieve a shorter processing time. Experiments were carried out comparing the SVFDT with the VFDT in 11 benchmark data stream datasets. This comparison assessed the trade-off between accuracy, memory, and processing time. Statistical analysis showed that the proposed algorithm obtained similar predictive performance and significantly reduced processing time and memory use. Thus, SVFDT is a suitable option for data stream mining with memory and time limitations, recommended as a weak learner in ensemble-based solutions.

A Flow feature detection framework for large-scale computational data based on incremental proper orthogonal decomposition and data mining

ABSTRACTA framework based on incremental proper orthogonal decomposition (iPOD) and data mining to perform large-scale computational data analysis is presented. It includes iPOD to incrementally reduce data dimensions and decouple dynamic flow structures in massive CFD data; data mining to classify and identify candidate global regions of interest (ROIs) for focused analysis; feature detection to capture key flow features and ultimate ROIs (UROIs); and targeted data storage and visualisation. Quantitative results show that iPOD is able to process large datasets that overwhelm the batch-POD, leading to 4–16× data reduction in the temporal domain. Data mining and feature detection algorithms, respectively, identify 50–70% of the spatial domain with high probability of flow feature occurrence and only 2–30% containing key flow features. The UROI and associated data can be selectively stored and visualised. In contrast to batch-POD, iPOD reduces memory usage by more than 10× and time by up to 75%.

PENERAPAN CITRA TERKOMPRESI PADA SEGMENTASI CITRA MENGGUNAKAN ALGORITMA K-MEANS

In the development of an image not only as a documentation of events. One area that requires image processing is in the field of medicine is radiology. In radiology there is a medical image required by doctors and researchers to be processed for patient analysis. One of the important problems in image processing and pattern recognition is image segmentation into homogeneous areas. Segmentation in medical images will result in a medical image with area boundaries that are important information for analysis. This research applies k-means algorithm to MRI (Magnetic Resonance Imaging) image segmentation. The input image used is the image of MRI (brain and breast) has gone through the compression stage. This compression process is done with the aim of reducing memory usage but the critical information content of MRI image is still maintained. The image of the segmentation result is evaluated through performance test using GCE, VOI, MSE, and PSNR parameters.

Integrated Signal Encoding and Analysis System for Actigraphy-based Long-term Monitoring of Periodic Limb Movements in Sleep.

There has been a boom in the development of wearable devices for wellness and healthcare applications. Numerous studies have been conducted on the utility of employing wearable devices for the long-term monitoring of biosignals. Despite their efficacy, the potential for practical implementation faces many hurdles such as memory usage, power consumption, denoising, and efficient data transmission. Of the many wearables being used, the actigraph has been a popular choice amongst experts for identifying motion abnormalities such as periodic leg movements (PLMs) in sleep and the activities of patients suffering from various medical illnesses. In this paper, we present an efficient pulse code modulation based, 3-bit, signal encoding technique, which when applied to long-term (6-8 hours), 16-bit sleep actigraphy signals, generates 3-bit encoded, accelerometry data with an average compression ratio of 92%, an average increase in the signal-to-noise (SNR) ratio by 20 dB and an average reduction of memory usage by 92%. The proposed technique also eliminates the need to apply filters for denoising, by retaining only characteristic signal information in the quantized version. The proposed technique, in general, could be applied to accelerometer-based wearables and has the potential to provide efficient memory and power usage in long-term monitoring applications.

Sparse Geometries Handling in Lattice Boltzmann Method Implementation for Graphic Processors

We describe a high-performance implementation of the lattice-Boltzmann method (LBM) for sparse geometries on graphic processors. In our implementation we cover the whole geometry with a uniform mesh of small tiles and carry out calculations for each tile independently with a proper data synchronization at tile edges. For this method we provide both the theoretical analysis of complexity and the results for real implementations for 2D and 3D geometries. Based on the theoretical model, we show that tiles offer significantly smaller bandwidth overhead than solutions based on indirect addressing. For 2-dimensional lattice arrangements a reduction of memory usage is also possible, though at the cost of diminished performance. We reached the performance of 682 MLUPS on GTX Titan (72\% of peak theoretical memory bandwidth) for D3Q19 lattice arrangement and double precision data.

Anisotropic Continuous-in-Scale Universal Multifractal Cascades: Simulation, Analysis and Correction Methods

Continuous-in-scale multifractal cascades has long been an attractive choice for mathematically modeling turbulent and turbulent-like geophysical fields. These fields are usually anisotropic as they are subject to both stratification and rotation, thereby questioning the isotropy assumption often made to model them. The self-affine and generalized scale invariance approaches to scaling are used here to introduce anisotropy in such models. These anisotropic simulations have (1) unresolved large-scale features and (2) statistics that deviate from the desired power-law scaling mainly in the small scales. The former issue is solved via nesting, whereas the latter is attempted to be overcome using singularity correction methods. While earlier studies have proposed isotropic correction methods, here they have been generalized to correct anisotropic simulations. These singularity corrections seem to improve the small-scale statistical properties of mildly anisotropic simulations; nesting, on the other hand, appears to enhance statistics over almost all scales even for strongly anisotropic simulations. Both the correction and nesting techniques lead to a reduction in computational time and memory usage suggesting that nested singularity-corrected cascades offer a better framework for quantitatively modeling the atmosphere, ocean, solid earth, and associated fields.

Reduced Energy and Memory Requirements by On-Board Behavior Classification for Animal-Borne Sensor Applications

The ability to study animal behavior is important in many fields of science, including behavioral ecology, conservation, and precision farming. These studies typically employ biotelemetry tags attached to animals that collect raw sensor data from tri-axial accelerometers and global positioning system modules. The lifespan of such tags is constrained by their power and memory usage, which are often limiting factors when performing behavioral studies for extended periods of time. This paper considers the power requirement and memory usage benefits of performing statistical behavior classification on the tag itself, as opposed to at a receiver station after raw data transmission. Experiments using specially designed low-power biotelemetry sensors demonstrated a 27-fold reduction in energy consumption and a 469-fold reduction in memory usage when classification was performed on the tag, rather than after raw data transmission. By performing on-board statistical behavior classification, both the power requirements and the memory usage are drastically reduced, thereby prolonging the lifespan of the tag.

SqueezedText: A Real-Time Scene Text Recognition by Binary Convolutional Encoder-Decoder Network

A new approach for real-time scene text recognition is proposed in this paper. A novel binary convolutional encoder-decoder network (B-CEDNet) together with a bidirectional recurrent neural network (Bi-RNN). The B-CEDNet is engaged as a visual front-end to provide elaborated character detection, and a back-end Bi-RNN performs character-level sequential correction and classification based on learned contextual knowledge. The front-end B-CEDNet can process multiple regions containing characters using a one-off forward operation, and is trained under binary constraints with significant compression. Hence it leads to both remarkable inference run-time speedup as well as memory usage reduction. With the elaborated character detection, the back-end Bi-RNN merely processes a low dimension feature sequence with category and spatial information of extracted characters for sequence correction and classification. By training with over 1,000,000 synthetic scene text images, the B-CEDNet achieves a recall rate of 0.86, precision of 0.88 and F-score of 0.87 on ICDAR-03 and ICDAR-13. With the correction and classification by Bi-RNN, the proposed real-time scene text recognition achieves state-of-the-art accuracy while only consumes less than 1-ms inference run-time. The flow processing flow is realized on GPU with a small network size of 1.01 MB for B-CEDNet and 3.23 MB for Bi-RNN, which is much faster and smaller than the existing solutions.

Sparse online feature maps

Online kernel methods suffer from computational and memory complexity in large-scale problems. Due to these drawbacks, budget online kernel learning and kernel approximation (low-dimensional feature map approximation) methods are widely used to speed up time and to reduce memory usage of kernel approaches. In this paper, orthogonal Gram-Schmidt explicit feature maps are applied to online kernel methods. The main advantage of these feature maps come from their orthogonality property. Utilization of these feature maps leads to mutually linearly independent dimensions of feature space, hence, reduce the redundancy in this space. These feature maps can be applied to single-pass online learning methods with l2- and l0-norm regularization to reduce the computational and memory complexity. In this paper, the proposed methods are named: 1) Online Feature Maps (OFEMs) and 2) Sparse Online Feature Maps (SOFEMs). These methods are examined for binary and multiclass single-label classification problems. Extensive experiments are compared with the results of other state-of-the-art methods on standard and real-world datasets. The experimental results show that OFEMs and SOFEMs outperform other methods in the literature.

A computationally efficient ductile damage model accounting for nucleation and micro-inertia at high triaxialities

The computational formulation of a micro-mechanical material model for the dynamic failure of ductile metals is presented in this paper. The statistical nature of porosity initiation is accounted for by introducing an arbitrary probability density function which describes the pores nucleation pressures. Each micropore within the representative volume element is modeled as a thick spherical shell made of plastically incompressible material. The treatment of porosity by a distribution of thick-walled spheres also allows for the inclusion of micro-inertia effects under conditions of shock and dynamic loading. The second order ordinary differential equation governing the microscopic porosity evolution is solved with a robust implicit procedure. A new Chebyshev collocation method is employed to approximate the porosity distribution and remapping is used to optimize memory usage. The adaptive approximation of the porosity distribution leads to a reduction of computational time and memory usage of up to two orders of magnitude. Moreover, the proposed model affords consistent performance: changing the nucleation pressure probability density function and/or the applied strain rate does not reduce accuracy or computational efficiency of the material model. The numerical performance of the model and algorithms presented is tested against three problems for high density tantalum: single void, one-dimensional uniaxial strain, and two-dimensional plate impact. The results using the integration and algorithmic advances suggest a significant improvement in computational efficiency and accuracy over previous treatments for dynamic loading conditions.

A Hybrid Boundary Element Model for Simulation and Optimization of Large Piezoelectric Micromachined Ultrasonic Transducer Arrays.

A hybrid boundary element model is proposed for the simulation of large piezoelectric micromachined ultrasonic transducer (PMUT) arrays in immersion. Multiphysics finite element method (FEM) simulation of a single-membrane structure is used to determine stiffness and piezoelectrically induced actuation loading of the membranes. To simulate the arrays of membranes in immersion, a boundary element method is employed, wherein membrane structures are modeled by a surface mesh that is coupled mechanically by mass, stiffness, and damping matrices, and acoustically by a mutual impedance matrix. A multilevel fast multipole algorithm speeds up computation time and reduces memory usage, enabling the simulation of thousands of membranes in a reasonable time. The model is validated with FEM for a small 3 3 matrix array for both square and circular membrane geometries. Two practical optimization examples of large PMUT arrays are demonstrated: membrane spacing of a 7 7 matrix array with circular membranes, and material choice and top electrode coverage of a 32-element linear array with 640 circular membranes. In addition, a simple analytical approach to electrode optimization based on normal mode theory is verified.

Content-based SMS Classification: Statistical Analysis for the Relationship between Features Size and Classification Performance

High dimensionality of the feature space is one of the difficulty that affect short message service(SMS) classification performance. Some studies used feature selection methods to pick up some features, while other studies used the full extracted features. In this work, we aim to analyse the relationship between features size and classification performance. For that, a classification performance comparison was carried outbetween ten features sizes selected by varies feature selection methods. The used methods were chi-square, Gini index and information gain (IG). Support vector machine was used as a classifier. Area Under the ROC (Receiver Operating Characteristics) Curve between true positive rate and false positive rate was used to measure the classification performance. We used the repeated measures ANOVA at p < 0.05 level to analyse the performance. Experimental results showed that IG method outperformed the other methods in all features sizes. The best result was with 50% of the extracted features. Furthermore, the results explicitly showed that using larger features size in the classification does not mean superior performance but sometimes leads to less classification performance. Therefore, feature selection step should be used. By reducing the used features for the classification, without degrading the classification performance, it means reducing memory usage and classification time.

Multithreaded implicitly dealiased convolutions

Implicit dealiasing is a method for computing in-place linear convolutions via fast Fourier transforms that decouples work memory from input data. It offers easier memory management and, for long one-dimensional input sequences, greater efficiency than conventional zero-padding. Furthermore, for convolutions of multidimensional data, the segregation of data and work buffers can be exploited to reduce memory usage and execution time significantly. This is accomplished by processing and discarding data as it is generated, allowing work memory to be reused, for greater data locality and performance. A multithreaded implementation of implicit dealiasing that accepts an arbitrary number of input and output vectors and a general multiplication operator is presented, along with an improved one-dimensional Hermitian convolution that avoids the loop dependency inherent in previous work. An alternate data format that can accommodate a Nyquist mode and enhance cache efficiency is also proposed.

R-SHT: A state history tree with R-Tree properties for analysis and visualization of highly parallel system traces

Abstract Understanding the behaviour of distributed computer systems with many threads and resources is a challenging task. Dynamic analysis tools such as tracers have been developed to assist programmers in debugging and optimizing the performance of such systems. However, complex systems can generate huge traces, with billions of events, which are hard to analyze manually. Trace visualization and analysis programs aim to solve this problem. Such software needs fast access to data, which a linear search through the trace cannot provide. Several programs have resorted to stateful analysis to rearrange data into more query friendly structures. In previous work, we suggested modifications to the State History Tree (SHT) data structure to correct its disk and memory usage. While the improved structure, eSHT, made near optimal disk usage and had reduced memory usage, we found that query performance, while twice as fast, exhibited scaling limitations. In this paper, we proposed a new structure using R-Tree techniques to improve query performance. We explain the hybrid scheme and algorithms used to optimize the structure to model the expected behaviour. Finally, we benchmark the data structure on highly parallel traces and on a demanding trace visualization use case. Our results show that the hybrid R-SHT structure retains the eSHT’s optimal disk usage properties while providing several orders of magnitude speed up to queries on highly parallel traces.

Memory-efficient 3D connected component labeling with parallel computing

Connected component labeling is a frequently used image processing task in many applications. Moreover, in recent years, the use of 3D image data has become widespread, for instance, in 3D X-ray computed tomography and magnetic resonance imaging. However, because ordinary labeling algorithms use a large amount of memory and 3D images are generally large, labeling 3D image data can cause memory shortages. Furthermore, labeling a large image is time-consuming. In this paper, we proposed new memory-efficient connected component labeling algorithm for 3D images with parallel computing. In this method, we accelerate the labeling process using parallel computing. In addition, we use a spans matrix and compressed label matrix to reduce memory usage. We also use an equivalence chain approach to speed up the calculation. Furthermore, the algorithm has two options for further processing performance or further memory savings. In the experiments on real examples, the proposed algorithm with the option for processing speed was faster and used less memory than the conventional label equivalence method. In contrast, with the proposed method using the memory-efficient option, we could further reduce memory from one-eighth to one-thirteenth that used by the label equivalence method while maintaining the same performance.

Hardware-and-memory-sharing architecture of deblocking filter for VP8 and H.264/AVC

In this paper, a hardware sharing architecture is proposed to support multi-standards--VP8 and H.264/AVC. To achieve a common architecture, the deblocking filters of VP8 and H.264/AVC are reorganized, and the lossless sharing architecture (RO-DBK) is obtained. For the lossy application such as low-resolution display devices, the modified coefficients to derive a highly sharing architecture (LC-DBK) are further adapted. The experimental results show that the PSNR only drops 0.36% on average for LC-DBK. These two proposed architectures save 74.2% and 80.2% arithmetic logic units, respectively. To further reduce memory usage, we propose a new VLSI architecture. In this VLSI architecture, the deblocking filter and the interpolation result of motion compensation and inverse transform share the same SRAM. To achieve this, a rearranged filtering order is also proposed. This design only uses 32 bytes transpose buffer without any SRAM usage. To reach real-time processing, a 3-staged pipeline scheduling is also proposed. The proposed design is implemented within 0.18 μm process. The working frequency is 100 MHz and the overall gate count is 3.9K.