Batch Strategy Research Articles

EALI: Energy-aware layer-level scheduling for convolutional neural network inference services on GPUs

The success of convolutional neural networks (CNNs) has made low-latency inference services on Graphic Processing Units (GPUs) a hot research topic. However, GPUs are hardware processors with high power consumption. To have the least energy consumption while meeting latency Service-Level-Objective (SLO), batching strategy and dynamic voltage frequency scaling (DVFS) are two important solutions. However, existing studies do not coordinate them and regard CNN as a black box, which makes inference services less energy-efficient. In this paper, we propose EALI, an energy-aware layer-level adaptive scheduling framework that is comprised of a power prediction model, a layer combination strategy, and an energy-aware layer-level scheduler. The power prediction model uses classic machine learning techniques to predict fine-grained layer-level power consumption. The layer combination strategy combines multiple layers into optimization units to lower scheduling overhead and complexity. The energy-aware layer-level scheduler adaptively coordinates batching strategy and layer-level DVFS according to workloads to minimize the energy consumption while meeting SLO. Our experimental results on NVIDIA Tesla M40 and V100 GPUs show that, compared to the state-of-the-art approaches, EALI decreases energy consumption by up to 36.24% while meeting SLO.

Training Physics‐Based Machine‐Learning Parameterizations With Gradient‐Free Ensemble Kalman Methods

AbstractMost machine learning applications in Earth system modeling currently rely on gradient‐based supervised learning. This imposes stringent constraints on the nature of the data used for training (typically, residual time tendencies are needed), and it complicates learning about the interactions between machine‐learned parameterizations and other components of an Earth system model. Approaching learning about process‐based parameterizations as an inverse problem resolves many of these issues, since it allows parameterizations to be trained with partial observations or statistics that directly relate to quantities of interest in long‐term climate projections. Here, we demonstrate the effectiveness of Kalman inversion methods in treating learning about parameterizations as an inverse problem. We consider two different algorithms: unscented and ensemble Kalman inversion. Both methods involve highly parallelizable forward model evaluations, converge exponentially fast, and do not require gradient computations. In addition, unscented Kalman inversion provides a measure of parameter uncertainty. We illustrate how training parameterizations can be posed as a regularized inverse problem and solved by ensemble Kalman methods through the calibration of an eddy‐diffusivity mass‐flux scheme for subgrid‐scale turbulence and convection, using data generated by large‐eddy simulations. We find the algorithms amenable to batching strategies, robust to noise and model failures, and efficient in the calibration of hybrid parameterizations that can include empirical closures and neural networks.

Evaluation of Lignocellulosic Wastewater Valorization with the Oleaginous Yeasts R. kratochvilovae EXF7516 and C. oleaginosum ATCC 20509

During the conversion of lignocellulose, phenolic wastewaters are generated. Therefore, researchers have investigated wastewater valorization processes in which these pollutants are converted to chemicals, i.e., lipids. However, wastewaters are lean feedstocks, so these valorization processes in research typically require the addition of large quantities of sugars and sterilization, which increase costs. This paper investigates a repeated batch fermentation strategy with Rhodotorula kratochvilovae EXF7516 and Cutaneotrichosporon oleaginosum ATCC 20509, without these requirements. The pollutant removal and its conversion to microbial oil were evaluated. Because of the presence of non-monomeric substrates, the ligninolytic enzyme activity was also investigated. The repeated batch fermentation strategy was successful, as more lipids accumulated every cycle, up to a total of 5.4 g/L (23% cell dry weight). In addition, the yeasts consumed up to 87% of monomeric substrates, i.e., sugars, aromatics, and organics acids, and up to 23% of non-monomeric substrates, i.e., partially degraded xylan, lignin, cellulose. Interestingly, lipid production was only observed during the harvest phase of each cycle, as the cells experienced stress, possibly due to oxygen limitation. This work presents the first results on the feasibility of valorizing non-sterilized lignocellulosic wastewater with R. kratochvilovae and C. oleaginosum using a cost-effective repeated batch strategy.

A Fine-Grain Batching-Based Task Allocation Algorithm for Spatial Crowdsourcing

Task allocation is a critical issue of spatial crowdsourcing. Although the batching strategy performs better than the real-time matching mode, it still has the following two drawbacks: (1) Because the granularity of the batch size set obtained by batching is too coarse, it will result in poor matching accuracy. However, roughly designing the batch size for all possible delays will result in a large computational overhead. (2) Ignoring non-stationary factors will lead to a change in optimal batch size that cannot be found as soon as possible. Therefore, this paper proposes a fine-grained, batching-based task allocation algorithm (FGBTA), considering non-stationary setting. In the batch method, the algorithm first uses variable step size to allow for fine-grained exploration within the predicted value given by the multi-armed bandit (MAB) algorithm and uses the results of pseudo-matching to calculate the batch utility. Then, the batch size with higher utility is selected, and the exact maximum weight matching algorithm is used to obtain the allocation result within the batch. In order to cope with the non-stationary changes, we use the sliding window (SW) method to retain the latest batch utility and discard the historical information that is too far away, so as to finally achieve refined batching and adapt to temporal changes. In addition, we also take into account the benefits of requesters, workers, and the platform. Experiments on real data and synthetic data show that this method can accomplish the task assignment of spatial crowdsourcing effectively and can adapt to the non-stationary setting as soon as possible. This paper mainly focuses on the spatial crowdsourcing task of ride-hailing.

Enhanced production of Bacopa saponins by repeated batch strategy in bioreactor.

Cultivation of cell suspension culture of Bacopa monnieri targeting the production of bacosides was explored in a 5-l stirred tank reactor using statistically optimized conditions. The bioreactor cultivation conditions were modified and this led to profuse biomass growth (2.81 ± 0.20g/l) and total bacosides (1.26 ± 0.23mg/g in cells and 0.60 ± 0.11mg/l in fermenter broth) production in 9days. The values of static volumetric mass transfer coefficient (kLa), dimensionless mixing time (Nm) were measured in the bioreactor. The culture grew efficiently and produced enhanced amount of bacoside A (5.59 ± 0.41mg/g total bacosides in cells and 3.12 ± 0.13mg/l in the fermenter broth) using one cycle of repeated batch strategy adopted in the bioreactor for 15days. The intracellular concentration of bacoside A3 (1.18 ± 0.11mg/g), bacopaside II (2.09 ± 0.35mg/g), bacopaside X (0.79 ± 0.17mg/g) and bacopasaponin C (2.24 ± 0.23mg/g) were significantly higher in repeated batch as compared to batch bioreactor cultivation. The yield of total bacosides in the fermenter broth was 5-times higher in repeated batch as compared to batch cultivation. This strategy can be helpful for the enhanced production of other valuable triterpenoid saponins.

Random Batch Particle Methods for the Homogeneous Landau Equation

We consider in this paper random batch particle methods for efficiently solving the homogeneous Landau equation in plasma physics. The methods are stochastic variations of the particle methods proposed by Carrillo et al. [J. Comput. Phys.: X 7: 100066, 2020] using the random batch strategy. The collisions only take place inside the small but randomly selected batches so that the computational cost is reduced to $O(N)$ per time step. Meanwhile, our methods can preserve the conservation of mass, momentum, energy and the decay of entropy. Several numerical examples are performed to validate our methods.

Improving the yield of graphene oxide-catalysed N-heterocyclization of amines through fed batch mode

The use of graphene oxide, a metal-free, heterogeneous carbocatalyst for a facile, efficient, and simple protocol for N-heterocyclization of aromatic amines with dihaloalkane to give azacycloalkanes and isolindolines in fed batch strategy was studied.

A Hybrid of Genetic Algorithm and Greedy Batch Strategy for Complex Job-Shop Scheduling Problem

A Hybrid of Genetic Algorithm and Greedy Batch Strategy for Complex Job-Shop Scheduling Problem

A multi-omics study to boost continuous bolaform sophorolipid production

Biodegradable and biobased surface active agents are renewable and environmentally friendly alternatives to petroleum derived or oleochemical surfactants. However, they are accompanied by relatively high production costs. In this study, the aim was to reduce the production costs for an innovative type of microbial biosurfactant: bolaform sophorolipids, produced by the yeast Starmerella bombicola ΔsbleΔat. A novel continuous retentostat set-up was performed whereby continuous broth microfiltration retained the biomass in the bioreactor while performing an in situ product separation of bolaform sophorolipids. Although a mean volumetric productivity of 0.56 g L−1 h−1 was achieved, it was not possible to maintain this productivity, which collapsed to almost 0 g L−1 h−1. Therefore, two process adaptations were evaluated, a sequential batch strategy and a phosphate limitation alleviation strategy. The sequential batch set-up restored the mean volumetric productivity to 0.66 g L−1 h−1 for an additional 132 h but was again followed by a productivity decline. A similar result was obtained with the phosphate limitation alleviation strategy where a mean volumetric productivity of 0.54 g L−1 h−1 was reached, but a productivity decline was also observed. Whole genome variant analysis uncovered no evidence for genomic variations for up to 1306 h of retentostat cultivation. Untargeted metabolomics analysis identified 8-hydroxyguanosine, a biomarker for oxidative RNA damage, as a key metabolite correlating with high bolaform sophorolipid productivity. This study showcases the application of a retentostat to increase bolaform sophorolipid productivity and lays the basis of a multi-omics platform for in depth investigation of microbial biosurfactant production with S. bombicola.

Dynamic Modeling of Tannase Production from Bacillus cereus: A Framework Simulation based on Fed Batch Strategy

Tannase enzyme is a metabolite of great interest in the industry. Typical examples of its use can be found in wines and beer production, beverage and juice fruits clarification and leather production. However, tannase production on an industrial scale is limited to the operation batch mode. For this reason, its production is low and severe limitations take the place of being carried on a large scale. To improve production, this research proposes an operation strategy based on Fed-batch mode. The kinetic constants were taken from the literature to simulate trends obtained through a Feed-batch mode of operation. One of the most important findings of this research focuses on increasing tannase production with found values of 0.380 U/g. The latter indicates that tannase production could be almost twice the concentration obtained with the traditional batch mode (0.1900 U/g). Results obtained in this research may be promising for the enzyme production industry. Using computational techniques, it is possible to identify an improvement without investing in excessive experimentation and resources.

To Batch or Not to Batch? Comparing Batching and Curriculum Learning Strategies across Tasks and Datasets

Many natural language processing architectures are greatly affected by seemingly small design decisions, such as batching and curriculum learning (how the training data are ordered during training). In order to better understand the impact of these decisions, we present a systematic analysis of different curriculum learning strategies and different batching strategies. We consider multiple datasets for three tasks: text classification, sentence and phrase similarity, and part-of-speech tagging. Our experiments demonstrate that certain curriculum learning and batching decisions do increase performance substantially for some tasks.

Smart resource allocation of concurrent execution of parallel applications

AbstractThread‐level parallelism (TLP) has been widely exploited to optimize computational resource usage in high‐performance systems. However, as many applications do not scale as the number of threads increase, resources will be wasted when the application executes with the maximum possible number of threads (i.e., the default execution) rather than fewer threads (thread throttling) that may use the resources more efficiently. Hence, instead of executing only one application with as many threads as possible, one can run more applications simultaneously by applying thread throttling to each one. The primary outcome of this strategy is a significant reduction in the total execution time and energy consumption when the system needs to execute a list of applications. Given that, we propose a smart resource allocation (SRA) for concurrent parallel application execution. It automatically finds the ideal degree of TLP for each application and guides the simultaneous parallel applications execution. When running 25 well‐known benchmarks on three multicore systems and comparing SRA to state‐of‐the‐art strategies (e.g., Batch, Equal policy, and Scalability), SRA improves the EDP by 87.4% over the Batch strategy; 75.5% over the Equal policy; and 38.8% over the scalability strategy.

Comparison of Spathaspora passalidarum and recombinant Saccharomyces cerevisiae for integration of first- and second-generation ethanol production.

First-generation ethanol (E1G) is based on the fermentation of sugars released from saccharine or starch sources, while second-generation ethanol (E2G) is focused on the fermentation of sugars released from lignocellulosic feedstocks. During the fractionation process to release sugars from hemicelluloses (mainly xylose), some inhibitor compounds are released hindering fermentation. Thus, the biggest challenge of using hemicellulosic hydrolysate is selecting strains and processes able to efficiently ferment xylose and tolerate inhibitors. With the aim of diluting inhibitors, sugarcane molasses (80% of sucrose content) can be mixed to hemicellulosic hydrolysate in an integrated E1G-E2G process. Cofermentations of xylose and sucrose were evaluated for the native xylose consumer Spathaspora passalidarum and a recombinant Saccharomyces cerevisiae strain. The industrial S. cerevisiae strain CAT-1 was modified to overexpress the XYL1, XYL2 and XKS1 genes and a mutant ([4-59Δ]HXT1) version of the low-affinity HXT1 permease, generating strain MP-C5H1. Although S. passalidarum showed better results for xylose fermentation, this yeast showed intracellular sucrose hydrolysis and low sucrose consumption in microaerobic conditions. Recombinant S. cerevisiae showed the best performance for cofermentation, and a batch strategy at high cell density in bioreactor achieved unprecedented results of ethanol yield, titer and volumetric productivity in E1G-E2G production process.

Learning graph-level representation from local-structural distribution with Graph Neural Networks

Graph Neural Networks has been proved to be successful in learning the latent vector representation for local structures. While applying this technique to some real-world tasks, such as property prediction for molecules, a graph-level representation generator should be required to integrate all these local-structural embeddings. Limited by varying graph scale and the absence of node order, existing graph-level representation generation approaches usually adopt rough ways, e.g., compressing all local-structural embeddings into a single vector, which cannot simultaneously retain both the local- and global-structural information in graph-level representations. To address this problem, this paper introduces a novel and more powerful approach to learn graph-level representation from local-structural distribution. Firstly, our approach employs a batch strategy to discretize the embeddings of local structures by their structural semantics, which can provide interactive information to approximately align the local-structural embeddings for varying graphs. According to Wasserstein metric, the aligned structural information is beneficial to capture the similarity among graphs. Then, our approach further integrates these aligned local-structural embeddings to construct a resolution-controllable structural histogram as the graph-level representation. Without over-compression, more local-structural information can be preserved. The experimental results show that our approach substantially outperformed the baselines on a range of real-world datasets in both graph classification and regression tasks.

A heuristic approach for the online order batching problem with multiple pickers

The Online Order Batching Problem with Multiple Pickers (OOBPMP) consists of optimizing the operations related to the picking process of orders in a warehouse, when the picking policy follows an order batching strategy. In this case, this variant of the well-known Order Batching Problem considers the existence of multiple workers in the warehouse and an online arrival of the orders. We study three different objective functions for the problem: minimizing the completion time, minimizing the picking time, and minimizing the differences in the workload among the pickers. We have identified and classified all previous works in the literature for the OOBPMP. Finally, we propose a multistart procedure hybridized with a Variable Neighborhood Descent metaheuristic to handle the problem. We test our proposal over well-known instances previously reported in the literature by empirically comparing the performance of our proposal with previous methods in the state of the art. The statistical tests corroborated the significance of the results obtained.

Improving Retail Warehouse Activity by Using Product Delivery Data

This paper presents a research study which is dedicated to the improvement in retail warehouse activity. This study aims to improve activity by identifying an efficient order picking strategy. (1) Background: The literature review shows the application of order picking strategies, but research related to their selection lacks an integrated approach. (2) Methods: The authors use the discrete event simulation method for the analysis of order picking strategies. The application of the discrete event simulation method enables various scenario tests in retail warehouses, allowing one to benchmark order picking strategies. By using the simulation model, experiments were designed to evaluate order picking strategies that are dependent on the delivery of the product distance variable. This research uses analysis of cost components and helps to identify the best possible order picking strategy to improve the overall warehouse performance. The authors benchmarked order picking strategies and presented constraints following product delivery data concerning their applications. (3) Results: The results presented show that the application of the order sorting strategy delivers 46.6% and the order batching strategy 6.7% lower costs compared to the single picking strategy. The results of the order batching strategy could be improved by 8.34% when the product clustering action is used. (4) Conclusions: The authors provide a theoretical framework which follows the application of order picking strategies using the product delivery data approach, which is the main scientific novelty of this paper. Recommendations are provided regarding the application of the proposed framework for the future improvement in retail warehouse activity.

Model-Based Evaluation of a Data-Driven Control Strategy: Application to Ibuprofen Crystallization

This work presents a methodology that relies on the application of the radial basis functions network (RBF)-based feedback control algorithms to a pharmaceutical crystallization process. Within the scope of the model-based evaluation of the proposed strategy, firstly strategies for the data treatment, data structure and the training methods reflecting the possible scenarios in the industry (Moving Window, Growing Window and Golden Batch strategies) were introduced. This was followed by the incorporation of such RBF strategies within a soft sensor application and a nonlinear predictive data-driven control application. The performance of the RBF control strategies was tested for the undisturbed cases as well as in the presence of disturbances in the process. The promising results from both RBF soft sensor control and the RBF predictive control demonstrated great potential of these techniques for the control of the crystallization process. In particular, both Moving Window and Golden Batch strategies performed the best results for an RBF soft sensor, and the Growing Window outperformed the remaining methodologies for predictive control.

Green methane production: Kinetic and mass transfer modeling in a batch process

One of the major issues in biomethanation studies, especially a batch strategy without mixing, is gaseous substrate mass transfer between gas and liquid phases. The strategy can be assumed as a simplified form of a stagnant underground gas reservoir. Hydrogen gas, as the limiting substrate, plays significant role in biomethanation. Being informed of hydrogen content diffused within the liquid phase for calculating percentage of active volume, help researcher to make a proper decision on bioreactor design or adjusting process parameters. For this purpose, a mass transfer modelling was developed which strengthened with a set of optimized biokinetic parameters. Parameter optimization was accomplished with the help of a predefined optimization algorithm and using a set of experimental data with the source of literature. Active volume calculation was successfully performed via the verified model and response surface methodology was served for maximizing it under variety of process conditions. It was found that the bioreactor height to width ratio significantly affected on active volume followed by pressure and temperature. In addition, working with a bioreactor with a circle cross section, in comparison with a square one, improved the maximum active volume up to 43% due to providing higher surface area for mass transfer. Sensitivity analysis verified the previous findings and revealed that higher pressures and temperatures linearly increased the active volume while increasing the bioreactor height to width ratio exponentially decreased the response. Furthermore, a wide bioreactor have potential to promote active volume up to 72% rather than a vertical one.

Mass Transfer Modeling For CO2 Removal from Environment with the Aim of Green Biomethanation and Methanogens Growth Optimization

Background and Objective: CO2 concentration, as the main greenhouse gas, is growing in atmosphere and many alternatives have been investigated to deal with it. However, harnessing with the aim of biomethanation seems to be more economic. Method: In this study a mass transfer modeling was conducted for a biomethanation process under a batch strategy aiming at maximizing liquid active volume. The accuracy of modeling results was assessed via comparing with experimental data and kinetic results under zero-dimension study. Then one-dimensional study was conducted in order to investigate biomass and hydrogen concentration profiles within liquid phase of the bioreactor and active volume calculation. Response surface method (RSM) was also served to investigate effect of temperature, pressure and as three main factors on active volume followed by response optimization. Findings: Model accuracy was confirmed by zero-dimension study. One-dimensional study was also revealed that biomass growth dispersion within liquid phase depends on hydrogen profile concentration on condition that both hydrogen and biomass diffusion coefficients were assumed to be equal. Their degree of magnification was 10-9 in standard conditions. RSM showed that the three studied factors significantly affected on bioreactor active volume. Meanwhile, pressure and temperature influenced the most, respectively. Discussion and Conclusion: A batch bioreactor with and high pressure and temperature met optimal conditions for biomethanation; however, process economy defines operational limitations.

Deep Reinforcement Learning for Task Assignment in Spatial Crowdsourcing and Sensing

With the continuous popularization of smart devices, Spatial Crowdsourcing(SC) came into being a novel paradigm of collecting smart sensing. As a core issue, the task assignment aims to arrange tasks to suitable workers for different objectives. However, most existing approaches matched tasks and workers under the premise of already known their number, while the arrival information of crowdsourcing participants cannot be understood in advance for their dynamically appearing. We fill this gap by proposing a Dynamic Task Assignment Framework based on Prediction and Adaptive Batching (DTAF-PAB). Specifically, we first predict the number of coming tasks and/or workers by a deep learning model Gated Recurrent Unit (GRU). Then, the adaptive batching strategies, based on Deep Q Network (DQN) and Double DQN (DDQN), are investigated to address the problem of adaptive and dynamic batching for task assignment. We design a deep reinforcement learning (DRL)-based algorithm to improve the overall benefit of the assignment. Finally, we use real-world datasets to evaluate the competitiveness of DTAF-PAB, and the experimental results show that the proposed framework is superior to other existing methods in terms of both predictive performance and crowdsourcing platform benefit.