Intermediate Data Storage Research Articles

Design and Evaluation of Real-Time Data Storage and Signal Processing in a Long-Range Distributed Acoustic Sensing (DAS) Using Cloud-Based Services.

In cloud-based Distributed Acoustic Sensing (DAS) sensor data management, we are confronted with two primary challenges. First, the development of efficient storage mechanisms capable of handling the enormous volume of data generated by these sensors poses a challenge. To solve this issue, we propose a method to address the issue of handling the large amount of data involved in DAS by designing and implementing a pipeline system to efficiently send the big data to DynamoDB in order to fully use the low latency of the DynamoDB data storage system for a benchmark DAS scheme for performing continuous monitoring over a 100 km range at a meter-scale spatial resolution. We employ the DynamoDB functionality of Amazon Web Services (AWS), which allows highly expandable storage capacity with latency of access of a few tens of milliseconds. The different stages of DAS data handling are performed in a pipeline, and the scheme is optimized for high overall throughput with reduced latency suitable for concurrent, real-time event extraction as well as the minimal storage of raw and intermediate data. In addition, the scalability of the DynamoDB-based data storage scheme is evaluated for linear and nonlinear variations of number of batches of access and a wide range of data sample sizes corresponding to sensing ranges of 1-110 km. The results show latencies of 40 ms per batch of access with low standard deviations of a few milliseconds, and latency per sample decreases for increasing the sample size, paving the way toward the development of scalable, cloud-based data storage services integrating additional post-processing for more precise feature extraction. The technique greatly simplifies DAS data handling in key application areas requiring continuous, large-scale measurement schemes. In addition, the processing of raw traces in a long-distance DAS for real-time monitoring requires the careful design of computational resources to guarantee requisite dynamic performance. Now, we will focus on the design of a system for the performance evaluation of cloud computing systems for diverse computations on DAS data. This system is aimed at unveiling valuable insights into performance metrics and operational efficiencies of computations on the data in the cloud, which will provide a deeper understanding of the system's performance, identify potential bottlenecks, and suggest areas for improvement. To achieve this, we employ the CloudSim framework. The analysis reveals that the virtual machine (VM) performance decreases significantly the processing time with more capable VMs, influenced by Processing Elements (PEs) and Million Instructions Per Second (MIPS). The results also reflect that, although a larger number of computations is required as the fiber length increases, with the subsequent increase in processing time, the overall speed of computation is still suitable for continuous real-time monitoring. We also see that VMs with lower performance in terms of processing speed and number of CPUs have more inconsistent processing times compared to those with higher performance, while not incurring significantly higher prices. Additionally, the impact of VM parameters on computation time is explored, highlighting the importance of resource optimization in the DAS system design for efficient performance. The study also observes a notable trend in processing time, showing a significant decrease for every additional 50,000 columns processed as the length of the fiber increases. This finding underscores the efficiency gains achieved with larger computational loads, indicating improved system performance and capacity utilization as the DAS system processes more extensive datasets.

Optimal Re-Materialization Strategies for Heterogeneous Chains: How to Train Deep Neural Networks with Limited Memory

Training in Feed Forward Deep Neural Networks is a memory-intensive operation which is usually performed on GPUs with limited memory capacities. This may force data scientists to limit the depth of the models or the resolution of the input data if data does not fit in the GPU memory. The re-materialization technique, whose idea comes from the checkpointing strategies developed in the Automatic Differentiation literature, allows data scientists to limit the memory requirements related to the storage of intermediate data (activations), at the cost of an increase in the computational cost. This paper introduces a new strategy of re-materialization of activations that significantly reduces memory usage. It consists in selecting which activations are saved and which activations are deleted during the forward phase, and then recomputing the deleted activations when they are needed during the backward phase. We propose an original computation model that combines two types of activation savings: either only storing the layer inputs, or recording the complete history of operations that produced the outputs. This paper focuses on the fully heterogeneous case, where the computation time and the memory requirement of each layer is different. We prove that finding the optimal solution is NP-hard and that classical techniques from Automatic Differentiation literature do not apply. Moreover, the classical assumption of memory persistence of materialized activations, used to simplify the search of optimal solutions, does not hold anymore. Thus, we propose a weak memory persistence property and provide a Dynamic Program to compute the optimal sequence of computations. This algorithm is made available through the Rotor software, a PyTorch plug-in dealing with any network consisting of a sequence of layers, each of them having an arbitrarily complex structure. Through extensive experiments, we show that our implementation consistently outperforms existing re-materialization approaches for a large class of networks, image sizes and batch sizes.

AoCStream: All-on-Chip CNN Accelerator with Stream-Based Line-Buffer Architecture and Accelerator-Aware Pruning.

Convolutional neural networks (CNNs) play a crucial role in many EdgeAI and TinyML applications, but their implementation usually requires external memory, which degrades the feasibility of such resource-hungry environments. To solve this problem, this paper proposes memory-reduction methods at the algorithm and architecture level, implementing a reasonable-performance CNN with the on-chip memory of a practical device. At the algorithm level, accelerator-aware pruning is adopted to reduce the weight memory amount. For activation memory reduction, a stream-based line-buffer architecture is proposed. In the proposed architecture, each layer is implemented by a dedicated block, and the layer blocks operate in a pipelined way. Each block has a line buffer to store a few rows of input data instead of a frame buffer to store the whole feature map, reducing intermediate data-storage size. The experimental results show that the object-detection CNNs of MobileNetV1/V2 and an SSDLite variant, widely used in TinyML applications, can be implemented even on a low-end FPGA without external memory.

I/O Performance Improvement of FHE Apriori with Striping File Layout Considering Storage of Intermediate Data

Fully homomorphic encryption (FHE) enables secret computations. Users can perform computation using data encrypted with FHE without decryption. Uploading private data without encryption to a public cloud has the risk of data leakage, which makes many users hesitant to utilize a public cloud. Uploading data encrypted with FHE avoids this risk, while still providing the computing power of the public cloud. In many cases, data are stored in HDDs because the data size increases significantly when FHE is used. One important data analysis is Apriori data mining. In this application, two files are accessed alternately, and this causes long-distance seeking on its HDD and low performance. In this paper, we propose a new striping layout with reservations for write areas. This method intentionally fragments files and arranges blocks to reduce the distance between blocks in a file and another file. It reserves the area for intermediate files of FHE Apriori. The performance of the proposed method was evaluated based on the I/O processing of a large FHE Apriori, and the results showed that the proposed method could improve performance by up to approximately 28%.

An Energy-Efficient Deep Convolutional Neural Network Training Accelerator for In Situ Personalization on Smart Devices

A scalable deep-learning accelerator supporting the training process is implemented for device personalization of deep convolutional neural networks (CNNs). It consists of three processor cores operating with distinct energy-efficient dataflow for different types of computation in CNN training. Unlike the previous works where they implement design techniques to exploit the same characteristics from the inference, we analyze major issues that occurred from training in a resource-constrained system to resolve the bottlenecks. A masking scheme in the propagation core reduces a massive amount of intermediate activation data storage. It eliminates frequent off-chip memory accesses for holding the generated activation data until the backward path. A disparate dataflow architecture is implemented for the weight gradient computation to enhance PE utilization while maximally reuse the input data. Furthermore, the modified weight update system enables an 8-bit fixed-point computing datapath. The processor is implemented in 65-nm CMOS technology and occupies 10.24 mm2 of the core area. It operates with the supply voltage from 0.63 to 1.0 V, and the computing engine runs in near-threshold voltage of 0.5 V. The chip consumes 40.7 mW at 50 MHz with the highest efficiency and achieves 47.4 $\mu \text{J}$ /epoch of training efficiency for the customized CNN model.

Computational Restructuring: Rethinking Image Compression Using Resistive Crossbar Arrays

Image compression is performed on billions of edge devices deployed in the Internet of Things (IoT). The bottleneck of the compression is the 2-D discrete cosine transform (2D DCT), which involves performing two matrix-matrix multiplications in series. Earlier studies have explored directly mapping the 2D DCT computation to emerging resistive crossbar arrays (RCAs), which promise to perform matrix-vector multiplication (MVM) with extremely small energy-delay product. The main drawback is that the series computation is inherently vulnerable to errors. In this article, we propose to fundamentally rethink how to perform image compression using RCAs. The key idea is to restructure the computation to natively match the properties of the underlying resistive hardware. This allows three of the main design steps within image compression (2D DCT, quantization, and zig-zag reordering) to be integrated into a single analog MVM operation. The integration is facilitated by the development of a 2D DCT reconstruction technique, a frequency spectrum optimization technique, and a quantization optimization technique. The techniques improve the robustness to errors, eliminates the storage of intermediate data, enables processing of small image blocks, facilitates the utilization of large-scale RCAs, and reduces the requirements on the expensive domain interfaces. Compared with the previous work, the experimental results demonstrate significant improvements in image quality while reducing power and latency with up to 62% and 21%, respectively.

Parallelized implementation of an explicit finite element method in many integrated core (MIC) architecture

Hardware accelerators are becoming increasingly important in boosting high performance computing systems. In this study, we develop a parallel explicit finite element (FE) analysis system based on a many integrated core (MIC) architecture for fast simulation of nonlinear dynamic problems of plate and shell structures. To minimize data transfer between heterogeneous architectures, parallel computation of the all explicit FE calculation is realized by developing a vectorized thread-level parallelism algorithm. The parallelism includes a novel dependency relationship link based method for efficiently solving parallel explicit shell element equations. A heterogeneous model is established to overlap data transfer and offloaded computation, and thus reduce the time required for large intermediate data storage in the actual engineering nonlinear problem simulation. Finally, a high performance nonlinear dynamic simulation system is developed. The simulations of benchmarks and engineering problems show that the parallel computing method proposed in this paper can give full play to the hardware performance of MIC architecture and effectively improve the computation efficiency of an explicit FE solution. For a bus body model containing approximately 3.8 million degrees of freedom, the computational speed is improved 17 times over CPU sequential computation, and the relative speedup grows with the increasing number of threads, the highest relative speedup exceeds 80.

A Comprehensive Study of MapReduce Over Lustre for Intermediate Data Placement and Shuffle Strategies on HPC Clusters

With high performance interconnects and parallel file systems, running MapReduce over modern High Performance Computing (HPC) clusters has attracted much attention due to its uniqueness of solving data analytics problems with a combination of Big Data and HPC technologies. Since the MapReduce architecture relies heavily on the availability of local storage media, the Lustre-based global storage in HPC clusters poses many new opportunities and challenges. In this paper, we perform a comprehensive study on different MapReduce over Lustre deployments and propose a novel high-performance design of YARN MapReduce on HPC clusters by utilizing Lustre as the additional storage provider for intermediate data. With a deployment architecture where both local disks and Lustre are utilized for intermediate data storage, we propose a novel priority directory selection scheme through which RDMA-enhanced MapReduce can choose the best intermediate storage during runtime by on-line profiling. Our results indicate that, we can achieve 44 percent performance benefit for shuffle-intensive workloads in leadership-class HPC systems. Our priority directory selection scheme can improve the job execution time by 63 percent over default MapReduce while executing multiple concurrent jobs. To the best of our knowledge, this is the first such comprehensive study for YARN MapReduce with Lustre and RDMA.

Porting HPC applications to the cloud: A multi-frontal solver case study

In this paper we argue that scientific applications traditionally considered as representing typical HPC workloads can be successfully and efficiently ported to a cloud infrastructure. We propose a porting methodology that enables parallelization of communication – and memory-intensive applications while achieving a good communication to computation ratio and a satisfactory performance in a cloud infrastructure. This methodology comprises several aspects: (1) task agglomeration heuristic enabling increasing granularity of tasks while ensuring they will fit in memory; (2) task scheduling heuristic increasing data locality; and (3) two-level storage architecture enabling in-memory storage of intermediate data. We implement this methodology in a scientific workflow system and use it to parallelize a multi-frontal solver for finite-element meshes, deploy it in a cloud, and execute it as a workflow. The results obtained from the experiments confirm that the proposed porting methodology leads to a significant reduction of communication costs and achievement of a satisfactory performance. We believe that these results constitute a valuable step toward a wider adoption of cloud infrastructures for computational science applications.

Improved CTT-SP Algorithm with Critical Path Method for Massive Data Storage in Scientific Workflow Systems

Scientific WorkFlows (SWFs) play significant roles in scientific research and engineering simulation, which are often data intensive and has complex data dependencies. The storage of massive intermediate datasets has great impacts on the performance and the quality of service of a SWF system, which has become a difficult and complex task. Through analyzing the cost transitive tournament shortest path (CTT-SP)-based algorithm proposed for the intermediate data storage in SWF systems, the disadvantage of the CTT-SP algorithm over its sensitivity to the main branches can be found. We proposed an improved CTT-SP algorithm based on the critical path method (CPM), aiming at reducing the sensitivity of the main branch and improving the performance of the algorithm. By adding the generation information of the datasets on the arrows, the data dependency graph of a SWF can be converted to an activity on arrow (AOA) net. Then, the critical path of the AOA net is used to be the main branch of the CTT-SP algorithm, which can reduce the impacts of the main branch and keep the quality of service of the SWF systems. Experiments are designed to test the impacts of the main branches and the performance of the improved CTT-SP algorithms. The results show the significant impacts of the main branches on the performance of the CTT-SP algorithm and the effectiveness of the CPM on improving the performance of the CTT-SP algorithm. Comparison results demonstrate the positiveness and effectiveness of the improved CTT-SP algorithm.

A Digital Liquid State Machine With Biologically Inspired Learning and Its Application to Speech Recognition.

This paper presents a bioinspired digital liquid-state machine (LSM) for low-power very-large-scale-integration (VLSI)-based machine learning applications. To the best of the authors' knowledge, this is the first work that employs a bioinspired spike-based learning algorithm for the LSM. With the proposed online learning, the LSM extracts information from input patterns on the fly without needing intermediate data storage as required in offline learning methods such as ridge regression. The proposed learning rule is local such that each synaptic weight update is based only upon the firing activities of the corresponding presynaptic and postsynaptic neurons without incurring global communications across the neural network. Compared with the backpropagation-based learning, the locality of computation in the proposed approach lends itself to efficient parallel VLSI implementation. We use subsets of the TI46 speech corpus to benchmark the bioinspired digital LSM. To reduce the complexity of the spiking neural network model without performance degradation for speech recognition, we study the impacts of synaptic models on the fading memory of the reservoir and hence the network performance. Moreover, we examine the tradeoffs between synaptic weight resolution, reservoir size, and recognition performance and present techniques to further reduce the overhead of hardware implementation. Our simulation results show that in terms of isolated word recognition evaluated using the TI46 speech corpus, the proposed digital LSM rivals the state-of-the-art hidden Markov-model-based recognizer Sphinx-4 and outperforms all other reported recognizers including the ones that are based upon the LSM or neural networks.

Supporting dynamic pipeline changes using Class-Based Object Versioning in Astro-WISE

Understanding the difference between data objects is a major problem especially in a scientific collaboration which allows scientists to collectively reuse data, modify and adapt scripts developed by their peers to process data while publishing the results to a centralized data store. Although data provenance has been significantly studied to address the origins of a data item, it does not however address changes made to the source code. Systems often appear as a large number of modules each containing hundreds of lines of code. It is, in general, not obvious which parts of source code contributed to the change in data object. The paper introduces the Class-Based Object Versioning framework, which overcomes some of the shortcomings of popular versioning systems (e.g. CVS, SVN) in maintaining data and code provenance information in scientific computing environments. The framework automatically identifies and captures useful fine-grained changes in the data and code of scripts that perform scientific experiments so that important information about intermediate stages (i.e. unrecorded changes in experiment parameters and procedures) can be identified and analyzed. The benefits of such a system include querying specific methods and code attributes for data items of interest, finding missing gaps of data lineage and implicit storage of intermediate data.

A data dependency based strategy for intermediate data storage in scientific cloud workflow systems

SUMMARYMany scientific workflows are data intensive where large volumes of intermediate data are generated during their execution. Some valuable intermediate data need to be stored for sharing or reuse. Traditionally, they are selectively stored according to the system storage capacity, determined manually. As doing science in the cloud has become popular nowadays, more intermediate data can be stored in scientific cloud workflows based on a pay‐for‐use model. In this paper, we build an intermediate data dependency graph (IDG) from the data provenance in scientific workflows. With the IDG, deleted intermediate data can be regenerated, and as such we develop a novel intermediate data storage strategy that can reduce the cost of scientific cloud workflow systems by automatically storing appropriate intermediate data sets with one cloud service provider. The strategy has significant research merits, i.e. it achieves a cost‐effective trade‐off of computation cost and storage cost and is not strongly impacted by the forecasting inaccuracy of data sets' usages. Meanwhile, the strategy also takes the users' tolerance of data accessing delay into consideration. We utilize Amazon's cost model and apply the strategy to general random as well as specific astrophysics pulsar searching scientific workflows for evaluation. The results show that our strategy can reduce the overall cost of scientific cloud workflow execution significantly. Copyright © 2010 John Wiley & Sons, Ltd.

Intermediate Holographic Data Storage System by Using Sequentially Superimposed Recording

We introduce a holographic data storage system for intermediating between small data sets and mass holographic data recording. It employs a holographic sequentially superimposed recording technique. We discuss a time scheduling technique for making uniform reconstruction of sequentially recorded holograms and we show experimental results. We also discuss the Bragg selectivity of sequentially recorded holograms. The maximum storage density of our system is estimated to be 224kbit/mm2. Our system is useful as an intermediate recording system before recording mass holographic data in a larger system.

An in-place architecture for the deblocking filter in H.264/AVC

This brief presents an in-place computing design for the deblocking filter used in H.264/AVC video coding standard. The proposed in-placed computing flow reuses intermediate data as soon as data is available. Thus, the intermediate data storage is reduced to only the four 4 /spl times/ 4 blocks instead of whole 16 /spl times/ 16 macroblock. The resulting design can achieve 100 MHz with only 13.41K gate count and support real-time deblocking operation of 2K /spl times/ 1K@30 Hz video application when clocked at 73.73 MHz by using 0.25-/spl mu/m CMOS technology.

The LONI Pipeline Processing Environment

The analysis of raw data in neuroimaging has become a computationally entrenched process with many intricate steps run on increasingly larger datasets. Many software packages exist that provide either complete analyses or specific steps in an analysis. These packages often possess diverse input and output requirements, utilize different file formats, run in particular environments, and have limited abilities with certain types of data. The combination of these packages to achieve more sensitive and accurate results has become a common tactic in brain mapping studies but requires much work to ensure valid interoperation between programs. The handling, organization, and storage of intermediate data can prove difficult as well. The LONI Pipeline Processing Environment is a simple, efficient, and distributed computing solution to these problems enabling software inclusion from different laboratories in different environments. It is used here to derive a T1-weighted MRI atlas of the human brain from 452 normal young adult subjects with fully automated processing. The LONI Pipeline Processing Environment’s parallel processing efficiency using an integrated client/server dataflow model was 80.9% when running the atlas generation pipeline from a PC client (Acer TravelMate 340T) on 48 dedicated server processors (Silicon Graphics Inc. Origin 3000). The environment was 97.5% efficient when the same analysis was run on eight dedicated processors.

Processing of multi-dimensional NMR data with the new software PROSA

The new program PROSA is an efficient implementation of the common data-processing steps for multi-dimensional NMR spectra. PROSA performs linear prediction, digital filtering, Fourier transformation, automatic phase correction, and baseline correction. High efficiency is achieved by avoiding disk storage of intermediate data and by the absence of any graphics display, which enables calculation in the batch mode and facilitates porting PROSA on a variety of different computer systems; including supercomputers. Furthermore, all time-consuming routines are completely vectorized. The elimination of a graphics display was made possible by the use of a new, reliable automatic phase-correction routine. CPU times for complete processing of a typical heteronuclear three-dimensional NMR data set of a protein vary between less than 1 min on a NEC SX3 supercomputer and 40 min on a Sun-4 computer system.

A new system for LEED intensity measurements using a real-time digital video processor

A new system for low-energy electron diffraction (LEED) intensity measurements has been developed using a video camera and digital processing of the video signal. Complete two-dimensional LEED patterns are digitized in real time with high resolution using a commercial video processor. Intensity-voltage (I-V) data on all beams in complex LEED patterns are collected simultaneously. A microcomputer analysis program automatically tracks the diffraction beams as a function of energy and calculates beam position, size, and integrated intensity, including a local background correction. Using a video tape recorder for intermediate data storage, a complete set of I-V curves can be collected in less than 100 s.

Optical reconstruction of NMR images

A method is presented for the instantaneous optical reconstruction of magnetic resonance images directly from free induction decay data. The method, which is holographic, depends on the identity between the FID (free induction decay) and the Fourier transform of the nuclear spin density with respect to the time integral of the magnetic field gradients. The method may be implemented using simple and inexpensive equipment, and in principle requires no intermediate data storage.

Three-dimensional systolic architecture for beamforming

A three-dimensional systolic architecture is described that efficiently performs passive broadband interpolation beam-forming for an array of sensors. The architecture transforms from time domain to beam space without intermediate data storage. The required convolution operations are implemented with linear sets of multiply/accumulate elements.