Search

Various scientific applications with different I/O characteristics are executed in HPC systems. However, underlying parallel file systems are unaware of these characteristics of applications, and using a single fixed file layout for all applications can degrade the performance of HPC systems. In this paper, we propose A2FL, an autonomous and adaptive file layout adjustment scheme that optimizes parallel file system configurations by analyzing the access pattern of the applications. The key steps of A2FL are as follows: (1) A2FL initially intercepts the I/O operations of the application, recording their access patterns in real-time. (2) The access patterns are then transformed into a graphical representation used for predicting I/O performance and providing adjustment recommendations. (3) A2FL autonomously adjusts the file layout based on the prediction results, delivering an optimal file layout within the parallel file system. Moreover, we propose A2FL-Compound which analyzes an access pattern by dividing it into smaller components to optimize the file layout in a fine-grained manner. Our evaluations demonstrate that A2FL significantly enhances I/O performance, with improvements of up to 65.9× compared to the default file layout.

Large community of high-energy physicists share their data all around world making it necessary to ship a large number of files over wide- area networks. Regional disk caches such as the Southern California Petabyte Scale Cache have been deployed to reduce the data access latency. We observe that about 94% of the requested data volume were served from this cache, without remote transfers, between Sep. 2022 and July 2023. In this paper, we show the predictability of the resource utilization by exploring the trends of recent cache usage. The time series based prediction is made with a machine learning approach and the prediction errors are small relative to the variation in the input data. This work would help understanding the characteristics of the resource utilization and plan for additional deployments of caches in the future.

Energy consumption in buildings accounts for a significant portion of the global energy use. Consequently, understanding building energy use is important. Data over the past decade show that the energy intensity (Joules/sqft) of commercial buildings has decreased. While some of the improvements (decrease in energy use) are easily measurable such as the use of more energy efficient lighting, impact of other modifications such as changes to the operation of the HVAC system or changes in the usage pattern of the building potentially due to external events are difficult to quantify. Simply comparing energy consumption prior and post change is not accurate as energy use is impacted by many factors including external weather conditions. In this paper, we present a case study to quantify the impact of external events on the energy consumption of a medium-sized office building. We adopt an approach based on counterfactual analysis. Towards this end, we first build two models based on Linear Regression and k-Nearest Neighbors to predict the daily energy use given different input features related to the weather. We determine the statistical features of the weather that are most predictive of energy use. We then use the models to determine a counterfactual baseline and thereby to accurately estimate the impact of the events. The results of the counterfactual analysis provide new insights on the impact of the events on energy consumption. The update to the building cooling system resulted in more energy savings than direct yearly comparison reveals. On the other hand, the tests of a MPC-based controller for the HVAC system saved less energy than determined by the direct yearly comparison. Finally, the results show that there no gains in terms of energy savings due to remote work during the COVID-19 pandemic. An increase in airflow setting in the HVAC system corroborates this finding and further validates the underlying model and the counterfactual analyses.

What makes you hold on to that old car? While the vast majority of household vehicles are still powered by conventional internal combustion engines, the progress of adopting emerging vehicle technologies will critically depend on how soon the existing vehicles are transacted out of the household fleet. Leveraging a nationally representative longitudinal data set, the Panel Study of Income Dynamics, this study examines how household decisions to dispose of or replace a given vehicle are: 1) influenced by the vehicle’s attributes, 2) mediated by households’ concurrent socio-demographic and economic attributes, and 3) triggered by key life cycle events. Coupled with a newly developed machine learning interpretation tool, TreeExplainer, we demonstrate an innovative use of machine learning models to augment traditional logit modeling to both generate behavioral insights and improve model performance. We find the two gradient-boosting-based methods, CatBoost and LightGBM, are the best performing machine learning models for this problem. The multinomial logistic model can achieve similar performance levels after its model specification is informed by TreeExplainer. Both machine learning and multinomial logit models suggest that while older vehicles are more likely to be disposed of or replaced than newer ones, such probability decreases as the vehicles serve the family longer. Pickup trucks and sport utility vehicles are less likely to be disposed of or replaced than cars, and leased vehicles are more likely to be transacted than owned vehicles. We find that married families, families with higher education levels, homeowners, and older families tend to keep their vehicles longer. Life events such as childbirth, residential relocation, and change of household composition and income are found to increase vehicle disposal and/or replacement. We provide additional insights on the timing of vehicle replacement or disposal, in particular, the presence of children and childbirth events are more strongly associated with vehicle replacement among younger parents.

Entry into parenthood is a major disruptive event to travel behavior, and gender gaps in mobility choices are often widened during parenthood. The exact timing of gender gap formation and their long-term effects on different subpopulations are less studied in the literature. Leveraging a longitudinal dataset from the 2018 WholeTraveler Study, this paper examines the effects of parenthood on a diverse set of short- to long-term outcomes related to the three hierarchical domains of mobility biography: mode choice, vehicle ownership, spatial mobility, and career decisions. The progress of the effects is evaluated over a sequential set of parenting stages and differentiated across three subpopulations. We find that individuals classified as “Have-it-alls”, who start their careers, partner up, and have children concurrently and early, significantly increase their car uses two years prior to childbirth (“nesting period”), and they then relocate to less transit-accessible areas and consequently reduce their reliance on public transportation while they have children in the household. In contrast, individuals categorized as “Couples”, who start careers and partnerships early but delay parenthood, and “Singles”, who postpone partnership and parenthood, have less pronounced changes in travel behavior throughout the parenting stages. The cohort-level effects are found to be driven primarily by women, whose career development is on average more negatively impacted by parenting events than men, regardless of their life course trajectory. Early career decisions made by women upon entering parenthood contribute to gender gaps in mid- to longer-term mobility decisions, signifying the importance of early intervention.

Scientific computing heavily relies on data shared by the community, especially in distributed data-intensive applications. This research focuses on predicting slow connections that create bottlenecks in distributed workflows. In this study, we analyze network traffic logs collected between January 2021 and August 2022 at the National Energy Research Scientific Computing Center (NERSC). Based on the observed patterns, we define a set of features primarily based on history for identifying low-performing data transfers. Typically, there are far fewer slow connections on well-maintained networks, which creates difficulty in learning to identify these abnormally slow connections from the normal ones. We devise several stratified sampling techniques to address the class-imbalance challenge and study how they affect the machine learning approaches. Our tests show that a relatively simple technique that undersamples the normal cases to balance the number of samples in two classes (normal and slow) is very effective for model training. This model predicts slow connections with an F1 score of 0.926.

Large-scale high performance computing (HPC) systems typically consist of many thousands of CPUs and storage units used by hundreds to thousands of users simultaneously. Applications from large numbers of users have diverse characteristics, such as varying computation, communication, memory, and I/O intensity. A good understanding of the performance characteristics of each user application is important for job scheduling and resource provisioning. Among these performance characteristics, I/O performance is becoming increasingly important as data sizes rapidly increase and large-scale applications, such as simulation and model training, are widely adopted. However, predicting I/O performance is difficult because I/O systems are shared among all users and involve many layers of software and hardware stack, including the application, network interconnect, operating system, file system, and storage devices. Furthermore, updates to these layers and changes in system management policy can significantly alter the I/O behavior of applications and the entire system. To improve the prediction of the I/O performance on HPC systems, we propose integrating information from several different system logs and developing a regression-based approach to predict the I/O performance. Our proposed scheme can dynamically select the most relevant features from the log entries using various feature selection algorithms and scoring functions, and can automatically select the regression algorithm with the best accuracy for the prediction task. The evaluation results show that our proposed scheme can predict the write performance with up to 90% prediction accuracy and the read performance with up to 99% prediction accuracy using the real logs from the Cori supercomputer system at NERSC.

The most common source of transformer failure is in the insulation, and the most prevalent warning signal for insulation weakness is partial discharge (PD). Locating the positions of these partial discharges would help repair the transformer to prevent failures. This work investigates algorithms that could be deployed to locate the position of a PD event using data from ultra-high frequency (UHF) sensors inside the transformer. These algorithms typically proceed in two steps: first determining the signal arrival time, and then locating the position based on time differences. This paper reviews available methods for each task and then propose new algorithms: a convolutional iterative filter with thresholding (CIFT) to determine the signal arrival time and a reference table of travel times to resolve the source location. The effectiveness of these algorithms are tested with a set of laboratory-triggered PD events and two sets of simulated PD events inside transformers in production use. Tests show the new approach provides more accurate locations than the best-known data analysis algorithms, and the difference is particularly large, 3.7X, when the signal sources are far from sensors.

In this work, we have analyzed the input/output (I/O) activities of Cori, which is a high-performance computing system at the National Energy Research Scientific Computing Center at Lawrence Berkeley National Laboratory. Our analysis results indicate that most users do not adjust storage configurations but rather use the default settings. In addition, owing to the interference from many applications running simultaneously, the performance varies based on the system status. To configure file systems autonomously in complex environments, we developed DCA-IO, a dynamic distributed file system configuration adjustment algorithm that utilizes the system log information to adjust storage configurations automatically. Our scheme aims to improve the application performance and avoid interference from other applications without user intervention. Moreover, DCA-IO uses the existing system logs and does not require code modifications, an additional library, or user intervention. To demonstrate the effectiveness of DCA-IO, we performed experiments using I/O kernels of real applications in both an isolated small-sized Lustre environment and Cori. Our experimental results shows that our scheme can improve the performance of HPC applications by up to 263% with the default Lustre configuration.

Distributed Acoustic Sensing (DAS) is an emerging sensing technology that records the strain-rate along fiber optic cables at high spatial and temporal resolution. This technique is becoming a popular tool in seismology, hydrology, and other subsurface monitoring applications. However, due to the large coverage (10’s of km) and high density of measurements (1m spacing at 100’s of Hz), a DAS installation could produce terabytes of data records per day. Because many DAS instruments are deployed in remote locations, this large data size poses significant challenges to its transfer and storage. In this paper, we explore lossless compression methods to reduce the storage requirement in both real-time and post-hoc scenarios. We propose a two-stage compression method to improve the compression ratio and compression speed. This two-stage compression method could reduce the storage requirement by 40%, which is 20% more than other lossless methods, such as ZSTD. We demonstrate that the compression method could complete its operation well before the DAS instrument needs to output the next file, making it suitable for real-time DAS acquisition. We also implement a parallel compression method for a post-hoc scenario and demonstrate that our method could effectively utilize a parallel computer. With 256 CPU cores, our parallel compression method achieves the speed of 26GB/second.

Data access throughput is one of the key performance metrics in scientific computing, particularly for distributed data-intensive applications. While there has been a body of studies focusing on elephant connections that consume a significant fraction of network bandwidth, this study focuses on predicting slow connections that create bottlenecks in distributed workflows. In this study, we analyze network traffic logs collected between January 2019 and May 2021 at National Energy Research Scientific Computing Center (NERSC). Based on the observed patterns from this data collection, we define a set of features to be used for identifying low-performing data transfers. Through extensive feature engineering and feature selection, we identify a number of new features to significantly enhance the prediction performance. With these new features, even the relatively simple decision tree model could predict slow connections with a F1 score as high as 0.945.

Motivated by extreme multi-label classification applications, we consider training deep learning models over sparse data in multi-GPU servers. The variance in the number of non-zero features across training batches and the intrinsic GPU heterogeneity combine to limit accuracy and increase the time to convergence. We address these challenges with Adaptive SGD, an adaptive elastic model averaging stochastic gradient descent algorithm for heterogeneous multi-GPUs that is characterized by dynamic scheduling, adaptive batch size scaling, and normalized model merging. Instead of statically partitioning batches to GPUs, batches are routed based on the relative processing speed. Batch size scaling assigns larger batches to the faster GPUs and smaller batches to the slower ones, with the goal to arrive at a steady state in which all the GPUs perform the same number of model updates. Normalized model merging computes optimal weights for every GPU based on the assigned batches such that the combined model achieves better accuracy. We show experimentally that Adaptive SGD outperforms four state-of-the-art solutions in time-to-accuracy and is scalable with the number of GPUs.

Neural networks are powerful solutions to many scientific applications; however, they usually require long model training time due to large training data sets or large model size. Research has been focused on developing numerical optimization algorithms and parallel processing to reduce the training time. In this work, we propose a multi-resolution strategy that can reduce the training time by training the model with the reduced-resolution data samples at the beginning and later switching to the original resolution data samples. This strategy is motivated by the observation that coarser versions of many applications can be solved faster than their denser counterparts, and the solution to a coarser problem could be used to initialize the solution to the denser problem. When applying the idea to neural network training, coarse data can have a similar effect on the learning curves at the early stage as the dense data but requires less time. Once the curves no longer improve significantly, our strategy switches to using the data in original resolution. The key in this process is the ability to generate multiple resolutions of a problem automatically, which could usually be done with scientific applications with spatial and temporal continuity. We use two real-world scientific applications, CosmoFlow and DeepCAM, to evaluate the proposed mixed-resolution training strategy. Our experiment results demonstrate that the proposed training strategy effectively reduces the end-to-end training time while achieving a comparable accuracy to that of the training only with the original data. While maintaining the same model accuracy, our multi-resolution training strategy reduces the end-to-end training time up to 30% and 23% for CosmoFlow and DeepCAM, respectively.

The applications being developed within the U.S. Exascale Computing Project (ECP) to run on imminent Exascale computers will generate scientific results with unprecedented fidelity and record turn-around time. Many of these codes are based on particle-mesh methods and use advanced algorithms, especially dynamic load-balancing and mesh-refinement, to achieve high performance on Exascale machines. Yet, as such algorithms improve parallel application efficiency, they raise new challenges for I/O logic due to their irregular and dynamic data distributions. Thus, while the enormous data rates of Exascale simulations already challenge existing file system write strategies, the need for efficient read and processing of generated data introduces additional constraints on the data layout strategies that can be used when writing data to secondary storage. We review these I/O challenges and introduce two online data layout reorganization approaches for achieving good tradeoffs between read and write performance. We demonstrate the benefits of using these two approaches for the ECP particle-in-cell simulation WarpX, which serves as a motif for a large class of important Exascale applications. We show that by understanding application I/O patterns and carefully designing data layouts we can increase read performance by more than 80 percent.

Many scientific applications have started using deep learning methods for their classification or regression problems. However, for data-intensive scientific applications, I/O performance can be the major performance bottleneck. In order to effectively solve important real-world problems using deep learning methods on High-Performance Computing (HPC) systems, it is essential to address the poor I/O performance issue in large-scale neural network training. In this paper, we propose an asynchronous I/O strategy that can be generally applied to deep learning applications. Our I/O strategy employs an I/O -dedicated thread per process, that performs I/O operations independently of the training progress. The I/O thread reads many training samples at once to reduce the total number of I/O operations per epoch. Given the fixed amount of training data, the fewer the I/O operations per epoch, the shorter the overall I/O time. The I/O operations are also overlapped with the computations using the double-buffering method. We evaluate our I/O strategy using two real-world scientific applications, CosmoFlow and Neuron-Inverter. Our experimental results demonstrate that the proposed I/O strategy significantly improves the scaling performance without affecting the regression performance.

High-performance computing (HPC) systems consist of thousands of compute nodes, storage systems and high-speed networks, providing multiple layers of I/O stack with high complexity. By adjusting the diverse configuration settings that HPC systems provide, the I/O performance of applications can be improved. However, it is challenging to identify the optimal configuration settings without a thorough knowledge of the system, as each of the different I/O characteristics of applications can be an important factor for parameter decision. In this paper, we use multiple machine learning approaches to perform an in-depth analysis on I/O behaviors of HPC applications and to search for the optimal configuration settings for jobs sharing similar I/O characteristics. Improved by maximum 0.07 R-squared score, our results in overall show that jobs run on the HPC systems can obtain the predicted I/O performance for different configuration parameters with a high accuracy, using the proposed machine learning-based prediction models.

Logistic regression has long been the gold standard for choice modeling in the transportation field. Despite the rising popularity of machine learning (ML), few is applied to predicting the household vehicle transactions. To address the research gap, this paper presents a first use case of ML application to predicting household vehicle transaction decisions by leveraging a newly processed national panel data set. Model performances are reported for four ML models and the traditional multinomial logit model (MNL). Instead of treating the gold standard and ML models as competitors, this paper tries to use ML tools to inform the MNL model building process. We find the two gradient boosting based methods, CatBoost and LightGBM, are the best performing ML models; and improving logistic models with SHAP interpretation tools can achieve similar performance levels to the best performing ML methods.

The widely-adopted practice is to train deep learning models with specialized hardware accelerators, e.g., GPUs or TPUs, due to their superior performance on linear algebra operations. However, this strategy does not employ effectively the extensive CPU and memory resources - which are used only for preprocessing, data transfer, and scheduling - available by default on the accelerated servers. In this paper, we study training algorithms for deep learning on heterogeneous CPU+GPU architectures. Our two-fold objective - maximize convergence rate and resource utilization simultaneously - makes the problem challenging. In order to allow for a principled exploration of the design space, we first introduce a generic deep learning framework that exploits the difference in computational power and memory hierarchy between CPU and GPU through asynchronous message passing. Based on insights gained through experimentation with the framework, we design two heterogeneous asynchronous stochastic gradient descent (SGD) algorithms. The first algorithm - CPU+GPU Hogbatch - combines small batches on CPU with large batches on GPU in order to maximize the utilization of both resources. However, this generates an unbalanced model update distribution which hinders the statistical convergence. The second algorithm - Adaptive Hogbatch - assigns batches with continuously evolving size based on the relative speed of CPU and GPU. This balances the model updates ratio at the expense of a customizable decrease in utilization. We show that the implementation of these algorithms in the proposed CPU+GPU framework achieves both faster convergence and higher resource utilization than TensorFlow on several real datasets.

To meet the exascale I/O requirements for the High-Performance Computing (HPC), a new I/O subsystem, Burst Buffer, based on solid state drives (SSD), has been developed. However, the diverse HPC workloads and the bursty I/O pattern cause severe data fragmentation that requires costly garbage collection (GC) and increases the number of bytes written to the SSD. To address this data fragmentation challenge, a new multi-stream feature has been developed for SSDs. In this work, we develop an I/O Separation scheme called BIOS to leverage this multi-stream feature to group the I/O streams based on the user IDs. We propose a stream-aware scheduling policy based on burst buffer pools in the workload manager, and integrate the BIOS with the workload manager to optimize the I/O separation scheme in burst buffer. We evaluate the proposed framework with a burst buffer I/O traces from Cori Supercomputer including a diverse set of applications. Experimental results show that the BIOS could improve the performance by 1.44x on average and reduce the Write Amplification Factor (WAF) by up to 1.20x. These demonstrate the potential benefits of the I/O separation scheme for solid state storage systems.

To understand behaviors of natural and man-made events, such as energy consumption of buildings, which accounts for 40% of energy uses in the US, we deploy automated monitoring devices to record periodic observations. However, such experimental and observation data often contains problems and irregularities that have to be cleaned up before analyses. Due to various conditions affecting sensor operations, the communication channels, recording steps, or the recording media, the recorded data might have missing values, errors, or anomalous values. An effective way to clean up these problems is to replace these missing values, errors and anomalous values with expected values, a process generally known as imputation. In this work, we survey commonly used missing value imputation techniques and compare their performance on a set of building monitoring data. To compare the different types of sensor measurements with widely varying characteristics, we use normalized root mean squared error (NRMSE) as the key metric for the effectiveness of the imputation methods. We additionally consider periodicity and run time when considering comparing methods. Through extensive testing, we find that for small gap sizes, up to 8 consecutive missing values, linear interpolation performs the best; for larger gaps stretching up to 48 consecutive missing values, K-nearest neighbors provides the most accurate imputations; for even larger gaps, more computational intensive methods, such as matrix factorization, achieve the smallest NRMSE. Additionally, we observe that these computationally intensive algorithms not only provide accurate imputations for large gaps, but are also more robust across all types of sensors.

Large community of high-energy physicists share their data all around world making it necessary to ship a large number of files over wide- area networks. Regional disk caches such as the Southern California Petabyte Scale Cache have been deployed to reduce the data access latency. We observe that about 94% of the requested data volume were served from this cache, without remote transfers, between Sep. 2022 and July 2023. In this paper, we show the predictability of the resource utilization by exploring the trends of recent cache usage. The time series based prediction is made with a machine learning approach and the prediction errors are small relative to the variation in the input data. This work would help understanding the characteristics of the resource utilization and plan for additional deployments of caches in the future.

The storage management system dCache acts as a disk cache for high-energy physics (HEP) data from the US ATLAS community. Since its disk capacity is considerably smaller than the total volume of ATLAS data, a heuristic is needed to determine what data should be kept on disks. An effective heuristic would be to keep the data files that are expected to be heavily accessed in the near future. Through a careful study of access statistics, we find a few most popular datasets are accessed much more frequently than others, even though these popular datasets change over time. If we could predict the near-term popularity of datasets, we could pin the most popular ones in the disk cache to prevent their accidental removal and guarantee their availability. To predict a dataset popularity, we present several methods for forecasting the number of times a dataset will be accessed in the next day. Test results show that these methods could predict the next-day access counts of popular datasets reliably. This observation is confirmed with dCache logs from two separate time ranges.

We present ADIOS 2, the latest version of the Adaptable Input Output (I/O) System. ADIOS 2 addresses scientific data management needs ranging from scalable I/O in supercomputers, to data analysis in personal computer and cloud systems. Version 2 introduces a unified application programming interface (API) that enables seamless data movement through files, wide-area-networks, and direct memory access, as well as high-level APIs for data analysis. The internal architecture provides a set of reusable and extendable components for managing data presentation and transport mechanisms for new applications. ADIOS 2 bindings are available in C++11, C, Fortran, Python, and Matlab and are currently used across different scientific communities. ADIOS 2 provides a communal framework to tackle data management challenges as we approach the exascale era of supercomputing.

With the growth of 5G wireless technologies and IoT, it become urgent to develop robust network security systems, such as intrusions detection systems (IDS) to keep the networks secure. These IDS systems need to detect unauthorized access and attacks in real-time. However, most of the modern IDS are built based on complex machine learning models that are time-consuming to train. In this work, we propose a methodology using the SHapley Additive exPlanations (SHAP) in combination with tree-based classifiers. SHAP can be used to select consistent and small feature subsets to reduce the execution time and improve classification accuracy. We demonstrate the proposed approach with the Aegean Wi-Fi Intrusion Dataset (AWID) dataset in a series of multi-class classification experiments. Among the four classes ("normal", "injection", "flooding"and "impersonation"), it is well-known that the class impersonation is hard to be classified accurately. Tests show that we can use about 10% of the initial feature set without reducing the overall prediction accuracy. With this reduced set of features, the training time could be reduced as much as a factor of four, while slightly improving the discriminating ability to identify impersonation instances. This study suggests that by reducing the number of features, the classification algorithms are able to focus on key trends that differentiates the "attacks"classes from the "normal"class. Using a reduces subset of features improves IDS's accuracy and performance. Also, SHAP dependence plots capture the relationship between individual features and the classification decision.