Search

It is widely accepted that cosmic rays are a plausible cause of DRAM errors in high-performance computing (HPC) systems, and various studies suggest that they could explain some aspects of the observed DRAM error behavior. However, this phenomenon is insufficiently studied in production environments. We analyze the correlations between cosmic rays and DRAM errors on two HPC clusters: a production supercomputer with server-class DDR3-1600 and a prototype with LPDDR3-1600 and no hardware error correction. Our error logs cover 2000 billion MB-hours for the MareNostrum 3 supercomputer and 135 million MB-hours for the Mont-Blanc prototype. Our analysis combines quantitative analysis, formal statistical methods and machine learning. We detect no indications that cosmic rays have any influence on the DRAM errors. To understand whether the findings are specific to systems under study, located at 100 meters above the sea level, the analysis should be repeated on other HPC clusters, especially the ones located on higher altitudes. Also, analysis can (and should) be applied to revisit and extend numerous previous studies which use cosmic rays as a hypothetical explanation for some aspects of the observed DRAM error behaviors., Comment: Accepted for publication in SBAC-PAD'24

Scaling to larger systems, with current levels of reliability, requires cost-effective methods to mitigate hardware failures. One of the main causes of hardware failure is an uncorrected error in memory, which terminates the current job and wastes all computation since the last checkpoint. This paper presents the first adaptive method for triggering uncorrected error mitigation. It uses a prediction approach that considers the likelihood of an uncorrected error and its current potential cost. The method is based on reinforcement learning, and the only user-defined parameters are the mitigation cost and whether the job can be restarted from a mitigation point. We evaluate our method using classical machine learning metrics together with a cost-benefit analysis, which compares the cost of mitigation actions with the benefits from mitigating some of the errors. On two years of production logs from the MareNostrum supercomputer, our method reduces lost compute time by 54% compared with no mitigation and is just 6% below the optimal Oracle method. All source code is open source., Comment: Published in HPDC'24

The Memory stress (Mess) framework provides a unified view of the memory system benchmarking, simulation and application profiling. The Mess benchmark provides a holistic and detailed memory system characterization. It is based on hundreds of measurements that are represented as a family of bandwidth--latency curves. The benchmark increases the coverage of all the previous tools and leads to new findings in the behavior of the actual and simulated memory systems. We deploy the Mess benchmark to characterize Intel, AMD, IBM, Fujitsu, Amazon and NVIDIA servers with DDR4, DDR5, HBM2 and HBM2E memory. The Mess memory simulator uses bandwidth--latency concept for the memory performance simulation. We integrate Mess with widely-used CPUs simulators enabling modeling of all high-end memory technologies. The Mess simulator is fast, easy to integrate and it closely matches the actual system performance. By design, it enables a quick adoption of new memory technologies in hardware simulators. Finally, the Mess application profiling positions the application in the bandwidth--latency space of the target memory system. This information can be correlated with other application runtime activities and the source code, leading to a better overall understanding of the application's behavior. The current Mess benchmark release covers all major CPU and GPU ISAs, x86, ARM, Power, RISC-V, and NVIDIA's PTX. We also release as open source the ZSim, gem5 and OpenPiton Metro-MPI integrated with the Mess simulator for DDR4, DDR5, Optane, HBM2, HBM2E and CXL memory expanders. The Mess application profiling is already integrated into a suite of production HPC performance analysis tools., Comment: 17 pages; accepted in MICRO-57 (2024), Best paper candidate

Domain-Specific Languages (DSLs) improve programmers productivity by decoupling problem descriptions from algorithmic implementations. However, DSLs for High-Performance Computing (HPC) have two additional critical requirements: performance and scalability. This paper presents the automated process of generating, from abstract mathematical specifications of Computational Fluid Dynamics (CFD) problems, optimised parallel codes that perform and scale as manually optimised ones. We consciously combine within Saiph, a DSL for solving CFD problems, low-level optimisations and parallelisation strategies, enabling high-performance single-core executions which effectively scale to multi-core and distributed environments. Our results demonstrate how high-level DSLs can offer competitive performance by transparently leveraging state-of-the-art HPC techniques., Comment: 30 pages, 18 figures. Postprint submitted to the Journal of Computational Science (Elsevier). Article updated with reviewers' comments, additional material in section 4.3 including figures and correction of typos

Task-based programming models are emerging as a promising alternative to make the most of multi-/many-core systems. These programming models rely on runtime systems, and their goal is to improve application performance by properly scheduling application tasks to cores. Additionally, these runtime systems offer policies to cope with application phases that lack in parallelism to fill all cores. However, these policies are usually static and favor either performance or energy efficiency. In this paper, we have extended a task-based runtime system with a lightweight monitoring and prediction infrastructure that dynamically predicts the optimal number of cores required for each application phase, thus improving both performance and energy efficiency. Through the execution of several benchmarks in multi-/many-core systems, we show that our prediction-based policies have competitive performance while improving energy efficiency when compared to state of the art policies., Comment: Postprint submitted and published at Euro-Par2020: International European Conference on Parallel and Distributed Computing (Springer) (https://link.springer.com/chapter/10.1007%2F978-3-030-57675-2_31)

In this work, we leverage ensemble learning as a tool for the creation of faster, smaller, and more accurate deep learning models. We demonstrate that we can jointly optimize for accuracy, inference time, and the number of parameters by combining DNN classifiers. To achieve this, we combine multiple ensemble strategies: bagging, boosting, and an ordered chain of classifiers. To reduce the number of DNN ensemble evaluations during the search, we propose EARN, an evolutionary approach that optimizes the ensemble according to three objectives regarding the constraints specified by the user. We run EARN on 10 image classification datasets with an initial pool of 32 state-of-the-art DCNN on both CPU and GPU platforms, and we generate models with speedups up to $7.60\times$, reductions of parameters by $10\times$, or increases in accuracy up to $6.01\%$ regarding the best DNN in the pool. In addition, our method generates models that are $5.6\times$ faster than the state-of-the-art methods for automatic model generation., Comment: 8 pages

Parallel task-based programming models, like OpenMP, allow application developers to easily create a parallel version of their sequential codes. The standard OpenMP 4.0 introduced the possibility of describing a set of data dependences per task that the runtime uses to order the tasks execution. This order is calculated using shared graphs, which are updated by all threads in exclusive access using synchronization mechanisms (locks) to ensure the dependence management correctness. The contention in the access to these structures becomes critical in many-core systems because several threads may be wasting computation resources waiting their turn. This paper proposes an asynchronous management of the runtime structures, like task dependence graphs, suitable for task-based programming model runtimes. In such organization, the threads request actions to the runtime instead of doing them directly. The requests are then handled by a distributed runtime manager (DDAST) which does not require dedicated resources. Instead, the manager uses the idle threads to modify the runtime structures. The paper also presents an implementation, analysis and performance evaluation of such runtime organization. The performance results show that the proposed asynchronous organization outperforms the speedup obtained by the original runtime for different benchmarks and different many-core architectures., Comment: 2020 Parallel Computing

In the image classification task, the most common approach is to resize all images in a dataset to a unique shape, while reducing their precision to a size which facilitates experimentation at scale. This practice has benefits from a computational perspective, but it entails negative side-effects on performance due to loss of information and image deformation. In this work we introduce the MAMe dataset, an image classification dataset with remarkable high resolution and variable shape properties. The goal of MAMe is to provide a tool for studying the impact of such properties in image classification, while motivating research in the field. The MAMe dataset contains thousands of artworks from three different museums, and proposes a classification task consisting on differentiating between 29 mediums (i.e. materials and techniques) supervised by art experts. After reviewing the singularity of MAMe in the context of current image classification tasks, a thorough description of the task is provided, together with dataset statistics. Experiments are conducted to evaluate the impact of using high resolution images, variable shape inputs and both properties at the same time. Results illustrate the positive impact in performance when using high resolution images, while highlighting the lack of solutions to exploit variable shapes. An additional experiment exposes the distinctiveness between the MAMe dataset and the prototypical ImageNet dataset. Finally, the baselines are inspected using explainability methods and expert knowledge, to gain insights on the challenges that remain ahead.

One of the major challenges in using extreme scale systems efficiently is to mitigate the impact of faults. Application-level checkpoint/restart (CR) methods provide the best trade-off between productivity, robustness, and performance. There are many solutions implementing CR at the application level. They all provide advanced I/O capabilities to minimize the overhead introduced by CR. Nevertheless, there is still room for improvement in terms of programmability and flexibility, because end-users must manually serialize and deserialize application state using low-level APIs, modify the flow of the application to consider restarts, or rewrite CR code whenever the backend library changes. In this work, we propose a set of compiler directives and clauses that allow users to specify CR operations in a simple way. Our approach supports the common CR features provided by all the CR libraries. However, it can also be extended to support advanced features that are only available in some CR libraries, such as differential checkpointing, the use of HDF5 format, and the possibility of using fault-tolerance-dedicated threads. The result of our evaluation revealed a high increase in programmability. On average, we reduced the number of lines of code by 71%, 94%, and 64% for FTI, SCR, and VeloC, respectively, and no additional overhead was perceived using our solution compared to using the backend libraries directly. Finally, portability is enhanced because our programming model allows the use of any backend library without changing any code.

Finding tumour genetic markers is essential to biomedicine due to their relevance for cancer detection and therapy development. In this paper, we explore a recently released dataset of chromosome rearrangements in 2,586 cancer patients, where different sorts of alterations have been detected. Using a Random Forest classifier, we evaluate the relevance of several features (some directly available in the original data, some engineered by us) related to chromosome rearrangements. This evaluation results in a set of potential tumour genetic markers, some of which are validated in the bibliography, while others are potentially novel.

High-resolution and variable-shape images have not yet been properly addressed by the AI community. The approach of down-sampling data often used with convolutional neural networks is sub-optimal for many tasks, and has too many drawbacks to be considered a sustainable alternative. In sight of the increasing importance of problems that can benefit from exploiting high-resolution (HR) and variable-shape, and with the goal of promoting research in that direction, we introduce a new family of datasets (MetH). The four proposed problems include two image classification, one image regression and one super resolution task. Each of these datasets contains thousands of art pieces captured by HR and variable-shape images, labeled by experts at the Metropolitan Museum of Art. We perform an analysis, which shows how the proposed tasks go well beyond current public alternatives in both pixel size and aspect ratio variance. At the same time, the performance obtained by popular architectures on these tasks shows that there is ample room for improvement. To wrap up the relevance of the contribution we review the fields, both in AI and high-performance computing, that could benefit from the proposed challenges., Comment: An improved and extended version of this paper has been published in arXiv:2007.13693 This version is now obsolete

The purpose of feature extraction on convolutional neural networks is to reuse deep representations learnt for a pre-trained model to solve a new, potentially unrelated problem. However, raw feature extraction from all layers is unfeasible given the massive size of these networks. Recently, a supervised method using complexity reduction was proposed, resulting in significant improvements in performance for transfer learning tasks. This approach first computes the discriminative power of features, and then discretises them using thresholds computed for the task. In this paper, we analyse the behaviour of these thresholds, with the purpose of finding a methodology for their estimation. After a comprehensive study, we find a very strong correlation between problem size and threshold value, with coefficient of determination above 90%. These results allow us to propose a unified model for threshold estimation, with potential application to transfer learning tasks., Comment: Presented in the 22nd International Conference of the Catalan Association for Artificial Intelligence (CCIA 19)

Continual learning based on data stream mining deals with ubiquitous sources of Big Data arriving at high-velocity and in real-time. Adaptive Random Forest ({\em ARF}) is a popular ensemble method used for continual learning due to its simplicity in combining adaptive leveraging bagging with fast random Hoeffding trees. While the default ARF size provides competitive accuracy, it is usually over-provisioned resulting in the use of additional classifiers that only contribute to increasing CPU and memory consumption with marginal impact in the overall accuracy. This paper presents Elastic Swap Random Forest ({\em ESRF}), a method for reducing the number of trees in the ARF ensemble while providing similar accuracy. {\em ESRF} extends {\em ARF} with two orthogonal components: 1) a swap component that splits learners into two sets based on their accuracy (only classifiers with the highest accuracy are used to make predictions); and 2) an elastic component for dynamically increasing or decreasing the number of classifiers in the ensemble. The experimental evaluation of {\em ESRF} and comparison with the original {\em ARF} shows how the two new components contribute to reducing the number of classifiers up to one third while providing almost the same accuracy, resulting in speed-ups in terms of per-sample execution time close to 3x.

Measuring the distance between concepts is an important field of study of Natural Language Processing, as it can be used to improve tasks related to the interpretation of those same concepts. WordNet, which includes a wide variety of concepts associated with words (i.e., synsets), is often used as a source for computing those distances. In this paper, we explore a distance for WordNet synsets based on visual features, instead of lexical ones. For this purpose, we extract the graphic features generated within a deep convolutional neural networks trained with ImageNet and use those features to generate a representative of each synset. Based on those representatives, we define a distance measure of synsets, which complements the traditional lexical distances. Finally, we propose some experiments to evaluate its performance and compare it with the current state-of-the-art.

The use of low-precision fixed-point arithmetic along with stochastic rounding has been proposed as a promising alternative to the commonly used 32-bit floating point arithmetic to enhance training neural networks training in terms of performance and energy efficiency. In the first part of this paper, the behaviour of the 12-bit fixed-point arithmetic when training a convolutional neural network with the CIFAR-10 dataset is analysed, showing that such arithmetic is not the most appropriate for the training phase. After that, the paper presents and evaluates, under the same conditions, alternative low-precision arithmetics, starting with the 12-bit floating-point arithmetic. These two representations are then leveraged using local scaling in order to increase accuracy and get closer to the baseline 32-bit floating-point arithmetic. Finally, the paper introduces a simplified model in which both the outputs and the gradients of the neural networks are constrained to power-of-two values, just using 7 bits for their representation. The evaluation demonstrates a minimal loss in accuracy for the proposed Power-of-Two neural network, avoiding the use of multiplications and divisions and thereby, significantly reducing the training time as well as the energy consumption and memory requirements during the training and inference phases., Comment: 16 pages, 9 figures and 4 tables

The current state-of-the-art for image annotation and image retrieval tasks is obtained through deep neural networks, which combine an image representation and a text representation into a shared embedding space. In this paper we evaluate the impact of using the Full-Network embedding in this setting, replacing the original image representation in a competitive multimodal embedding generation scheme. Unlike the one-layer image embeddings typically used by most approaches, the Full-Network embedding provides a multi-scale representation of images, which results in richer characterizations. To measure the influence of the Full-Network embedding, we evaluate its performance on three different datasets, and compare the results with the original multimodal embedding generation scheme when using a one-layer image embedding, and with the rest of the state-of-the-art. Results for image annotation and image retrieval tasks indicate that the Full-Network embedding is consistently superior to the one-layer embedding. These results motivate the integration of the Full-Network embedding on any multimodal embedding generation scheme, something feasible thanks to the flexibility of the approach., Comment: In 2nd Workshop on Semantic Deep Learning (SemDeep-2) at the 12th International Conference on Computational Semantics (IWCS) 2017

Patterns stored within pre-trained deep neural networks compose large and powerful descriptive languages that can be used for many different purposes. Typically, deep network representations are implemented within vector embedding spaces, which enables the use of traditional machine learning algorithms on top of them. In this short paper we propose the construction of a graph embedding space instead, introducing a methodology to transform the knowledge coded within a deep convolutional network into a topological space (i.e. a network). We outline how such graph can hold data instances, data features, relations between instances and features, and relations among features. Finally, we introduce some preliminary experiments to illustrate how the resultant graph embedding space can be exploited through graph analytics algorithms., Comment: Accepted at the 2nd Workshop on Semantic Deep Learning (SemDeep-2)

Transfer learning for feature extraction can be used to exploit deep representations in contexts where there is very few training data, where there are limited computational resources, or when tuning the hyper-parameters needed for training is not an option. While previous contributions to feature extraction propose embeddings based on a single layer of the network, in this paper we propose a full-network embedding which successfully integrates convolutional and fully connected features, coming from all layers of a deep convolutional neural network. To do so, the embedding normalizes features in the context of the problem, and discretizes their values to reduce noise and regularize the embedding space. Significantly, this also reduces the computational cost of processing the resultant representations. The proposed method is shown to outperform single layer embeddings on several image classification tasks, while also being more robust to the choice of the pre-trained model used for obtaining the initial features. The performance gap in classification accuracy between thoroughly tuned solutions and the full-network embedding is also reduced, which makes of the proposed approach a competitive solution for a large set of applications.

While cluster computing frameworks are continuously evolving to provide real-time data analysis capabilities, Apache Spark has managed to be at the forefront of big data analytics for being a unified framework for both, batch and stream data processing. There is also a renewed interest is Near Data Computing (NDC) due to technological advancement in the last decade. However, it is not known if NDC architectures can improve the performance of big data processing frameworks such as Apache Spark. In this position paper, we hypothesize in favour of NDC architecture comprising programmable logic based hybrid 2D integrated processing-in-memory and in-storage processing for Apache Spark, by extensive profiling of Apache Spark based workloads on Ivy Bridge Server., Comment: position paper

We introduce a community detection algorithm (Fluid Communities) based on the idea of fluids interacting in an environment, expanding and contracting as a result of that interaction. Fluid Communities is based on the propagation methodology, which represents the state-of-the-art in terms of computational cost and scalability. While being highly efficient, Fluid Communities is able to find communities in synthetic graphs with an accuracy close to the current best alternatives. Additionally, Fluid Communities is the first propagation-based algorithm capable of identifying a variable number of communities in network. To illustrate the relevance of the algorithm, we evaluate the diversity of the communities found by Fluid Communities, and find them to be significantly different from the ones found by alternative methods., Comment: Accepted at the 6th International Conference on Complex Networks and Their Applications

Deep neural networks are representation learning techniques. During training, a deep net is capable of generating a descriptive language of unprecedented size and detail in machine learning. Extracting the descriptive language coded within a trained CNN model (in the case of image data), and reusing it for other purposes is a field of interest, as it provides access to the visual descriptors previously learnt by the CNN after processing millions of images, without requiring an expensive training phase. Contributions to this field (commonly known as feature representation transfer or transfer learning) have been purely empirical so far, extracting all CNN features from a single layer close to the output and testing their performance by feeding them to a classifier. This approach has provided consistent results, although its relevance is limited to classification tasks. In a completely different approach, in this paper we statistically measure the discriminative power of every single feature found within a deep CNN, when used for characterizing every class of 11 datasets. We seek to provide new insights into the behavior of CNN features, particularly the ones from convolutional layers, as this can be relevant for their application to knowledge representation and reasoning. Our results confirm that low and middle level features may behave differently to high level features, but only under certain conditions. We find that all CNN features can be used for knowledge representation purposes both by their presence or by their absence, doubling the information a single CNN feature may provide. We also study how much noise these features may include, and propose a thresholding approach to discard most of it. All these insights have a direct application to the generation of CNN embedding spaces., Comment: Published in the Journal of Artificial Intelligence Research (JAIR), Special Track on Deep Learning, Knowledge Representation, and Reasoning

The hyperlink prediction task, that of proposing new links between webpages, can be used to improve search engines, expand the visibility of web pages, and increase the connectivity and navigability of the web. Hyperlink prediction is typically performed on webgraphs composed by thousands or millions of vertices, where on average each webpage contains less than fifty links. Algorithms processing graphs so large and sparse require to be both scalable and precise, a challenging combination. Similarity-based algorithms are among the most scalable solutions within the link prediction field, due to their parallel nature and computational simplicity. These algorithms independently explore the nearby topological features of every missing link from the graph in order to determine its likelihood. Unfortunately, the precision of similarity-based algorithms is limited, which has prevented their broad application so far. In this work we explore the performance of similarity-based algorithms for the particular problem of hyperlink prediction on large webgraphs, and propose a novel method which assumes the existence of hierarchical properties. We evaluate this new approach on several webgraphs and compare its performance with that of the current best similarity-based algorithms. Its remarkable performance leads us to argue on the applicability of the proposal, identifying several use cases of hyperlink prediction. We also describes the approach we took for the computation of large-scale graphs from the perspective of high-performance computing, providing details on the implementation and parallelization of code., Comment: Submitted to Transactions on Internet Technology journal

Link prediction, the problem of identifying missing links among a set of inter-related data entities, is a popular field of research due to its application to graph-like domains. Producing consistent evaluations of the performance of the many link prediction algorithms being proposed can be challenging due to variable graph properties, such as size and density. In this paper we first discuss traditional data mining solutions which are applicable to link prediction evaluation, arguing about their capacity for producing faithful and useful evaluations. We also introduce an innovative modification to a traditional evaluation methodology with the goal of adapting it to the problem of evaluating link prediction algorithms when applied to large graphs, by tackling the problem of class imbalance. We empirically evaluate the proposed methodology and, building on these findings, make a case for its importance on the evaluation of large-scale graph processing., Comment: Submitted to New Generation Computing, 15 pages, 4 tables, 4 figures

While cluster computing frameworks are continuously evolving to provide real-time data analysis capabilities, Apache Spark has managed to be at the forefront of big data analytics for being a unified framework for both, batch and stream data processing. However, recent studies on micro-architectural characterization of in-memory data analytics are limited to only batch processing workloads. We compare micro-architectural performance of batch processing and stream processing workloads in Apache Spark using hardware performance counters on a dual socket server. In our evaluation experiments, we have found that batch processing are stream processing workloads have similar micro-architectural characteristics and are bounded by the latency of frequent data access to DRAM. For data accesses we have found that simultaneous multi-threading is effective in hiding the data latencies. We have also observed that (i) data locality on NUMA nodes can improve the performance by 10% on average and(ii) disabling next-line L1-D prefetchers can reduce the execution time by up-to 14\% and (iii) multiple small executors can provide up-to 36\% speedup over single large executor.

Sheer increase in volume of data over the last decade has triggered research in cluster computing frameworks that enable web enterprises to extract big insights from big data. While Apache Spark is gaining popularity for exhibiting superior scale-out performance on the commodity machines, the impact of data volume on the performance of Spark based data analytics in scale-up configuration is not well understood. We present a deep-dive analysis of Spark based applications on a large scale-up server machine. Our analysis reveals that Spark based data analytics are DRAM bound and do not benefit by using more than 12 cores for an executor. By enlarging input data size, application performance degrades significantly due to substantial increase in wait time during I/O operations and garbage collection, despite 10\% better instruction retirement rate (due to lower L1 cache misses and higher core utilization). We match memory behaviour with the garbage collector to improve performance of applications between 1.6x to 3x., Comment: accepted to 6th International Workshop on Big Data Benchmarks, Performance Optimization and Emerging Hardware (BpoE-6) held in conjunction with VLDB 2015. arXiv admin note: text overlap with arXiv:1506.07742

In last decade, data analytics have rapidly progressed from traditional disk-based processing to modern in-memory processing. However, little effort has been devoted at enhancing performance at micro-architecture level. This paper characterizes the performance of in-memory data analytics using Apache Spark framework. We use a single node NUMA machine and identify the bottlenecks hampering the scalability of workloads. We also quantify the inefficiencies at micro-architecture level for various data analysis workloads. Through empirical evaluation, we show that spark workloads do not scale linearly beyond twelve threads, due to work time inflation and thread level load imbalance. Further, at the micro-architecture level, we observe memory bound latency to be the major cause of work time inflation., Comment: Accepted to The 5th IEEE International Conference on Big Data and Cloud Computing (BDCloud 2015)