Enhancing the performance of a microarray gridding algorithm via GPU computing techniques

An automated method for gridding and clustering-based segmentation of cDNA microarray images

The rapid growth of language models, particularly in code generation, requires substantial computational resources, raising concerns about energy consumption and environmental impact. Optimizing language models inference resource utilization is crucial, and Small Language Models (SLMs) offer a promising solution to reduce resource demands. Our goal is to analyze the impact of deep learning serving configurations, defined as combinations of runtime engines and execution providers, on resource utilization, in terms of energy consumption, execution time, and computing-resource utilization from the point of view of software engineers conducting inference in the context of code generation SLMs. We conducted a technology-oriented, multi-stage experimental pipeline using twelve code generation SLMs to investigate energy consumption, execution time, and computing-resource utilization across the configurations. Significant differences emerged across configurations. CUDA execution provider configurations outperformed CPU execution provider configurations in both energy consumption and execution time. Among the configurations, TORCH paired with CUDA demonstrated the greatest energy efficiency, achieving energy savings from 37.99% up to 89.16% compared to other serving configurations. Similarly, optimized runtime engines like ONNX with the CPU execution provider achieved from 8.98% up to 72.04% energy savings within CPU-based configurations. Also, TORCH paired with CUDA exhibited efficient computing-resource utilization. Serving configuration choice significantly impacts resource utilization. While further research is needed, we recommend the above configurations best suited to software engineers’ requirements for enhancing serving resource utilization efficiency. Editor’s note: Open Science material was validated by the Journal of Systems and Software Open Science Board . • Actionable guidelines to improve resource utilization of deep learning serving. • Measured impact of deep learning serving configurations on resource utilization. • A reproducible pipeline and analysis of deep learning serving configurations.

Task offloading is crucial for the efficient management of vehicular wireless networks, requiring access to real-time information on network status and available computational resources. However, because vehicles are mobile network nodes with variable speeds and limited computational resources for near-real-time task offloading, task offloading schemes may incur high latency, rendering this goal practically infeasible. To address this issue, this paper presents a trajectory-prediction-based pre-offloading decision (TPPD) algorithm that analyzes historical vehicle trajectories to predict future coordinates, thereby enabling the allocation of computational resources in advance. We first use a Long Short-Term Memory (LSTM) network to predict vehicle trajectories with error correction. Then, based on the task requirements and predicted trajectories, a dynamic resource allocation algorithm using a Double Deep Q-Network (DDQN) is designed to enable the multi-access edge server to minimize task processing delay while ensuring effective utilization of available computational resources. Numerical results verify the effectiveness of the proposed approach, demonstrating that, compared to traditional real-time task offloading strategies, the proposed TPPD algorithm significantly reduces task processing delay while improving resource utilization.

Microarray images, which allow the analysis of hybridization expressions of genes, have been one of the most widely used high-throughput technologies with many different applications. Accurate and automatic microarray image analysis is very important since researchers trust the information provided in these experiments and construct conclusions based on the results produced by the software responsible in analyzing the hybridized arrays. Every microarray image contains thousands of spots, so how to do the spots finding in microarray images accurately and automatically is very meaningful. There are always some problems, such as rotation and distortion, in a microarray image caused by mechanical errors and/or optical errors in the system. All these problems will hinder "doi"ng analysis automatically. Early research has worked out several algorithms to deal with the rotation problem, but those algorithms can not handle microarray images with distortions. In this paper, we propose a robust spots finding method to deal with both rotation and/or distortion in microarray images. The proposed method provides automatic gridding and can handle a microarray image with different type of rotation (global or sub-array rotation) and optical distortions.

Ability to work in a computation-limited and computation-variant environment described as "computational awareness" is a desirable feature of a realtime motion estimation system. In this paper, a new multi-path computation-aware algorithm is proposed. One-pass scheme presented in the prior work by Chen et al., considers only one candidate point which is the median of neighboring motion vectors (MVs). However, when computational resources are abundant, it is possible to investigate more search paths around points from the prediction set and get a significant improvement both in terms of quality and utilization of available search points (SPs). Also other enhancements are introduced regarding the way of selection of the starting SP and the search strategy, and allocation of resources. These enhancements lead to higher PSNR and a better utilization of computational resources.

BackgroundHigh-throughput Next-Generation Sequencing (NGS) techniques are advancing genomics and molecular biology research. This technology generates substantially large data which puts up a major challenge to the scientists for an efficient, cost and time effective solution to analyse such data. Further, for the different types of NGS data, there are certain common challenging steps involved in analysing those data. Spliced alignment is one such fundamental step in NGS data analysis which is extremely computational intensive as well as time consuming. There exists serious problem even with the most widely used spliced alignment tools. TopHat is one such widely used spliced alignment tools which although supports multithreading, does not efficiently utilize computational resources in terms of CPU utilization and memory. Here we have introduced PVT (Pipelined Version of TopHat) where we take up a modular approach by breaking TopHat’s serial execution into a pipeline of multiple stages, thereby increasing the degree of parallelization and computational resource utilization. Thus we address the discrepancies in TopHat so as to analyze large NGS data efficiently.ResultsWe analysed the SRA dataset (SRX026839 and SRX026838) consisting of single end reads and SRA data SRR1027730 consisting of paired-end reads. We used TopHat v2.0.8 to analyse these datasets and noted the CPU usage, memory footprint and execution time during spliced alignment. With this basic information, we designed PVT, a pipelined version of TopHat that removes the redundant computational steps during ‘spliced alignment’ and breaks the job into a pipeline of multiple stages (each comprising of different step(s)) to improve its resource utilization, thus reducing the execution time.ConclusionsPVT provides an improvement over TopHat for spliced alignment of NGS data analysis. PVT thus resulted in the reduction of the execution time to ~23% for the single end read dataset. Further, PVT designed for paired end reads showed an improved performance of ~41% over TopHat (for the chosen data) with respect to execution time. Moreover we propose PVT-Cloud which implements PVT pipeline in cloud computing system.

Microarray technology has become a power tool in the field of bioinformatics. It is used to measure gene expression levels and similar to any other image capturing processes is prone to noise. There are different kinds of noise, during preparation, hybridization and scanning in microarray images which usually are modeled by Gaussian noise. Since introduction of wavelets in 1970s, many more forms and extensions of this transform have been developed and used, such as stationary wavelet transform (SWT), complex wavelet transform (CWT), curvelet transform (CURV) and contourlet transform (CNT). By developing of more sparse transforms, it is important to have a perspective of how efficient the transforms are in different applications, such as microarray image analysis. In this paper, we compare the efficiency of common sparse transforms including ordinary discrete wavelet transform (DWT), SWT, CWT, CURV, CNT, Contourlet-SD decomposition, steerable pyramid (STP) and shearlet transform (SHR) for microarray image denoising. Therefore after converting microarray image into x-let transform, BayesShrink method, soft and hard thresholding are used to perform denoising of these images. Both local and general thresholds are calculated for each subband in order to evaluate the effect of incorporating intrascale dependency on top of sparsity property in statistical modeling of x-let's coefficients. Our simulation results show that CWT and SHR outperforms the others when using global thresholding and SWT is the preferred transform when using local thresholding. Although STP and SHR have better performance for some criteria like structural similarity (SSIM) index, but CWT is faster.

To exploit the benefits of throughput-optimized processors such as GPUs, applications need to be redesigned to achieve performance and efficiency. In this work, we present techniques to speed up statistical timing analysis on throughput processors. We draw upon advancements in improving the efficiency of Monte Carlo based statistical static timing analysis (MC SSTA) using techniques to reduce the sample size or smart sampling techniques. An efficient smart sampling technique, Stratification + Hybrid Quasi Monte Carlo (SH-QMC), is implemented on a GPU based on NVIDIA CUDA architecture. We show that although this application is based on MC analysis with straightforward parallelism available, achieving performance and efficiency on the GPU requires exposing more parallelism and finding locality in computations. This is in contrast with random sampling based algorithms which are inefficient in terms of sample size but can keep resources utilized on a GPU. We show that SH-QMC implemented on a Multi GPU is twice as fast as a single STA on a CPU for benchmark circuits considered. In terms of an efficiency metric, which measures the ability to convert a reduction in sample size to a corresponding reduction in runtime w.r.t a random sampling approach, we achieve 73.9% efficiency with the proposed approaches compared to 4.3% for an implementation involving performing computations on smart samples in parallel. Another contribution of the paper is a critical graph analysis technique to improve the efficiency of Monte Carlo based SSTA, leading to 2--9X further speedup.

Segmentation is a routine and crucial procedure for the treatment of brain tumors. Deep learning based brain tumor segmentation methods have achieved promising performance in recent years. However, to pursue high segmentation accuracy, most of them require too much memory and computation resources. Motivated by a recently proposed partially reversible U-Net architecture that pays more attention to memory footprint, we further present a novel Memory-Efficient Cascade 3D U-Net (MECU-Net) for brain tumor segmentation in this work, which can achieve comparable segmentation accuracy with less memory and computation consumption. More specifically, MECU-Net utilizes fewer down-sampling channels to reduce the utilization of memory and computation resources. To make up the accuracy loss, MECU-Net employs multi-scale feature fusion module to enhance the feature representation capability. Additionally, a light-weight cascade model, which resolves the problem of small target segmentation accuracy caused by model compression to some extent, is further introduced into the segmentation network. Finally, edge loss and weighted dice loss are combined to refine the brain tumor segmentation results. Experiment results on BraTS 2019 validation set illuminate that MECU-Net can achieve average Dice coefficients of 0.902, 0.824 and 0.777 on the whole tumor, tumor core and enhancing tumor, respectively.

In this paper, an industrial grade adaptive control scheme is proposed for a micro-grid integrated dual active bridge driven battery management system (DIBMS). A benchmark industrial grade adaptive control scheme depends on two factors namely robustness and computational resource utilization when such controllers are implemented over processors. The mathematical model of DIBMS system is nonlinear, thus for desired response, non-linear controllers based on sliding mode variable structure control theory suits it well for the state regulation problem of DIBMS, however such controllers utilize high computational resources when practically implemented over processors. Keeping in view the above performance indices, this paper proposes an industrial grade computationally efficient and finite time adaptive robust convergent control for DIBMS system. A proportional integral (PI) scheme is used as central control unit and Hebbian algorithm with double integration of the state error is introduced for online tuning the gains of central control unit. The robustness and computational resource efficiency of the proposed control paradigm is validated using a laboratory scale test bench through TI Launchpad (TMS320F28379D). The superiority of the proposed AI based PI control paradigm is compared with classical PI, integer order sliding mode control (SMC), and fractional order SMC (FOSMC) in terms of computational resource utilization and robustness under all test conditions.

Bead encoding is a key problem central to all multiplexed bead-based microarrays. Most existing bead-based microarrays require complicated and costly bead fabrication and/or sophisticated bead decoding hardware to achieve high multiplexing. Recently, the development of spatial bead encoding techniques have opened up the possibility of using pattern matching and image processing to develop highly multiplexed, high-throughput bead-based arrays. In this thesis, we improve the existing spatial bead encoding scheme, and develop a pipeline of computational methods, which allows for automated spatial bead encoding. Six novel computational methods, that automate the proposed bead encoding scheme, are developed in this thesis. The proposed scheme improves the previously reported spatial encoding schemes by making them better suited for use in real world scenarios. In the spatial bead encoding scheme, a sequential bead deposition method is used to capture the identities of the beads in bright or dark field images of the array. While, a second fluorescence image of the array is used to quantify the target analyte concentrations associated with the beads. By aligning the two images using the patterns formed by the beads, both the bead identities (i.e. the target analyte associate with the bead) and the analyte concentrations are decoded. The six computational methods developed in this thesis, are grouped into three categories: bright and dark field image processing, fluorescence image processing, and bead pattern matching. These methods are developed such that they require minimal parameter tuning, and are able to deal with noisy images acquired in uncontrolled environments As part of the bright and dark field image processing we have developed two novel methods: a fully automatic method for detecting the underlying micro-well grid structure in the images, and an unsupervised learning based method to classify the micro-wells as either empty or containing a bead. We show the ability of the proposed methods to deal with extreme amounts of noise and distortions using multiple datasets. While, microarray image gridding is a well-studied problem in the context of fluorescence images of planar microarrays, to the best of our knowledge no methods have yet been developed for processing either bright or dark field images of bead-based microarrays. Unlike in bright field images, where all the beads are visible, in fluorescence images only those beads that are expressed in the sample are visible. Therefore, in fluorescence image analysis the first step is to detect the beads in the image. To this end, we have developed an unsupervised, simultaneous bead detection and segmentation method which uses the coherence in shape and size of the beads to circumvent the need for parameter optimization. The main challenge in this segmentation is the large dynamic range across which the beads are expressed. We have also developed a probabilistic non-linear gridding method that attempts to establish the micro-well grid structure from the detected beads. While similar gridding techniques have been developed in the past, we make two contributions in this regard: first, we put the gridding method in a probabilistic framework, and second, we show that in certain cases it is impossible to ascertain if the correct grid has been detected without using additional information about the physical properties of the array. The final step in the proposed bead encoding scheme is to establishing one-to-one correspondences between the beads detected in the bright field and fluorescence images. We have developed two methods for this matching. When the micro-well grid structure in the fluorescence image is successfully established, the grid information can be used to reduce the matching problem to a binary grid alignment problem. A grid matching method that is very fast and can handle a large number of beads (>50,000) has been developed for this purpose. Finally, to perform matching when the fluorescence image grid is not available, a novel point pattern matching method that is affine invariant and robust to small non-linear distortions has been developed. The point pattern matching method addresses the problem of matching large point sets in the presence of large amounts of outliers and occlusion. The proposed method is evaluated using several real and simulated datasets.

<p>To design green clouds with optimal energy and resource consumption, the host consolidation process became the most prominent cloud management policy today. In case of a consolidation, the underloaded host's VMs are shifted to the other hosts to control the energy extravagance at a data center. The consolidation process is considered an NP-hard problem, as it depends on several computational resources (i.e., CPU, RAM, Bandwidth) and migration constraints. Although many former cloud researchers were focused on designing the consolidation process, the total consolidation process is based on single computational resource (i.e., CPU) utilization only. In a single resource-based consolidation model, only the selected resource utilization value is considered in decision-making and the other resource utilization values are simply discarded. The problems identified in single resource-based consolidation process are: (i) Ignoring the other prominent resources in consolidation (ii) unnecessary consolidation of hosts (iii) aggressive migrations and (iv) energy extravagance. To overcome the limitations involved with the single resource-based consolidation process, this study proposed the “Multi Resources and Constraints-based Consolidation Model (MRCCM)". The MRCCM is designed to solve the energy extravagance issues and aggressive migrations in host consolidation process. In MRCCM, multiple resources and resource constraints are considered while host consolidation for best target host selection and efficient resource management. Cloudsim toolkit with 800 HP-ProLiant host models and EC2 VM models are configured for experimental analysis. Proposed MRCCM model with MMT host selection method recorded the total 11570 migrations, 1853 shut_downs and 111.84 Kwh of energy consumption.  Experimental results are proven that the MRCCM controlled the energy extravagance, unnecessary consolidations, and aggressive migrations at a considerable rate when compared to the other prominent consolidation models.</p><p><br></p>

Automation of production processes of enterprise today has enormous scales. The spread of automated control systems of technological processes (ACSTP) and their integration with corporate information systems have led to increase in proportion of successfully implemented malicious impacts on ACSTP. Occurrence of new types of attacks and increase in number of successfully implemented attacks cause the need to form protection systems of industrial systems. Designing protection system of ACSTP is multi-stage process, starting with collection of data on object of protection, its main components, information assets, links with other systems, and identification of relevant information security threats to the system. In this article analysis of structure of typical automated control system of technological process is carried out. The most important components of ACSTP from the point of view of information security are identified. The article defines the need to form project of protection system of ACSTP on the basis of monitoring data of events that occur in ACSTP. Main classes of system events from the point of view of information security are defined. The most typical information security threats for this type of automated systems are considered. Neutralization of groups of threats relevant to ACSTP is carried out by means of forming effective protection system. The typical composition of data protection system of ACSTP is defined. According to results of analysis, the correspondence between security events of ACSTP, current threat groups and typical means of information protection is determined. Choosing the composition of information security tools is complex task that requires a lot of time. To achieve greater efficiency in its solving, it is proposed to use decision support system and system of artificial intelligence. Authors developed functional model of selection of protection system's project using intelligent method of decision support with neural network in it.

In this paper, an interval observer-based fault diagnosis mechanism and an interval dynamic event-triggered robust fault-tolerant control strategy are proposed for uncertain nonlinear switched systems. Firstly, the interval observer design problem is transformed into a Sylvester equation solution problem by introducing the coordinate transformation method in the absence of actuator faults, and the corresponding interval observer design method is given. Secondly, based on the interval observer output, upper and lower bound estimates of the system output are constructed, and fault diagnosis is realized by monitoring whether the system output exceeds the preset bounds. In addition, an improved interval dynamic event triggered mechanism is proposed to optimize the utilization of communication and computational resources and reduce redundant information transmission. The mechanism incorporates interval estimation information and real-time residuals, which can effectively extend the triggering time interval while improving the fault detection performance. Based on the interval fault diagnosis results, a fault-tolerant control strategy based on the interval dynamic event triggered mechanism is proposed, and the conditions for asymptotic stabilization of the closed-loop system are given. Finally, a nonlinear model of a morphing aircraft system with variable wing curvature is used to verify the validity of the designed scheme. DOI: 10.61416/ceai.v28i1.9516

As the amount of re-sequencing genome data grows, minimizing the execution time of an analysis is required. For this purpose, recent computing systems have been adopting both high-performance coprocessors and host processors. However, there are few applications that efficiently utilize these heterogeneous computing resources. This problem equally refers to the work of single nucleotide polymorphism (SNP) detection, which is one of the bottlenecks in genome data processing. In this paper, we propose a method for speeding up an SNP detection by enhancing the utilization of heterogeneous computing resources often used in recent high-performance computing systems. Through the measurement of workload in the detection procedure, we divide the SNP detection into several task groups suitable for each computing resource. These task groups are scheduled using a window overlapping method. As a result, we improved upon the speedup achieved by previous open source applications by a magnitude of 10.