Given the exponential data growth and the recent focus on understanding high-dimensional "in-motion" data, fundamental machine learning tools, such as Principal Component Analysis (PCA), require computation-efficient streaming algorithms that operate near-real-time. Despite the different streaming PCA flavors, there is no algorithm that provably recovers the principal components in the same precision regime as the batch PCA algorithm does, while maintaining low-latency and high-throughput processing. This work, introduces a novel temporal accumulate / retract learning framework for streaming PCA. We consider the accumulate / retract framework implementation of several competitive PCA algorithms with proven theoretical advantages. We benchmark the improved PCA algorithms on real-world streams (i.e. bank transactions fraud detection) and prove their low-latency (millisecond level) and high-throughput (thousands events/second) processing guarantees.

With the evolution of data acquisition methods, our ability to collect real time data has increased. This requires the development of real-time analytics, using the most recent data to generate valuable insights. One example is customer profiling, where we want to identify groups of similar clients who were active recently, and improve the quality of the suggestions. Traditional clustering algorithms perform well on finite datasets, but their execution is often not compatible with real-time requirements, especially for rapid changing trends. In this context, we propose a novel approach for the definition of incremental clustering algorithms to work within real-time constraints, in an online fashion, while preserving accuracy. We show the general applicability of the framework by employing this method to three different clustering algorithms. We compare the experimental results between traditional and online approaches evaluating accuracy and computational cost. The results show that algorithms executed in our framework are comparable to their offline implementation in terms of accuracy and with a high gain in execution time, up to three orders of magnitude on average.

Big Data applications are increasingly moving from batch-oriented execution models to stream-based models that enable them to extract valuable insights close to real-time. To support this model, an essential part of the streaming processing pipeline is data ingestion, i.e., the collection of data from various sources (sensors, NoSQL stores, filesystems, etc.) and their delivery for processing. Data ingestion needs to support high throughput, low latency and must scale to a large number of both data producers and consumers. Since the overall performance of the whole stream processing pipeline is limited by that of the ingestion phase, it is critical to satisfy these performance goals. However, state-of-art data ingestion systems such as Apache Kafka build on static stream partitioning and offset-based record access, trading performance for design simplicity. In this paper we propose KerA, a data ingestion framework that alleviate the limitations of state-of-art thanks to a dynamic partitioning scheme and to lightweight indexing, thereby improving throughput, latency and scalability. Experimental evaluations show that KerA outperforms Kafka up to 4x for ingestion throughput and up to 5x for the overall stream processing throughput. Furthermore, they show that KerA is capable of delivering data fast enough to saturate the big data engine acting as the consumer.

Nowadays, data sources, such as IoT devices, financial markets, and online services, continuously generate large amounts of data. Such data is usually generated at high frequencies and is typically described by non-stationary distributions. Querying these data sources brings new challenges for machine learning algorithms, which now need to be considered from the perspective of an evolving stream and not a static dataset. Under such scenarios, where data flows continuously, the challenge is how to transform the vast amount of data into information and knowledge, and how to adapt to data changes (i.e. drifts) and accumulate experience over time to support online decision-making. In this paper, we introduce STARLORD, a novel incremental computation method and system acting on data streams and capable of achieving low-latency (millisecond level) and high-throughput (thousands events/second/core) when learning from data streams. Moreover, the approach is able to adapt to data drifts and accumulate experience over time, and to use such knowledge to improve future learning and prediction performance, with resource usage guarantees. This is proven by our preliminary experiments where we built-in the framework in an open source stream engine (i.e. Apache Flink).

Big Data applications are rapidly moving from a batch-oriented execution model to a streaming execution model in order to extract value from the data in real-time. However, processing live data alone is often not enough: in many cases, such applications need to combine the live data with previously archived data to increase the quality of the extracted insights. Current streaming-oriented runtimes and middlewares are not flexible enough to deal with this trend, as they address ingestion (collection and pre-processing of data streams) and persistent storage (archival of intermediate results) using separate services. This separation often leads to I/O redundancy (e.g., write data twice to disk or transfer data twice over the network) and interference (e.g., I/O bottlenecks when collecting data streams and writing archival data simultaneously). In this position paper, we argue for a unified ingestion and storage architecture for streaming data that addresses the aforementioned challenge. We identify a set of constraints and benefits for such a unified model, while highlighting the important architectural aspects required to implement it in real life. Based on these aspects, we briefly sketch our plan for future work that develops the position defended in this paper.

International audience

Summary The emergence of cloud computing has brought the opportunity to use large‐scale compute infrastructures for a broader and broader spectrum of applications and users. As the cloud paradigm gets attractive for the ‘elasticity’ in resource usage and associated costs (the users only pay for resources actually used), cloud applications still suffer from the high latencies and low performance of cloud storage services. As Big Data analysis on clouds becomes more and more relevant in many application areas, enabling high‐throughput massive data processing on cloud data becomes a critical issue, as it impacts the overall application performance. In this paper, we address this challenge at the level of cloud storage. We introduce a concurrency‐optimized data storage system (called TomusBlobs), which federates the virtual disks associated to the Virtual Machines running the application code on the cloud. We demonstrate the performance benefits of our solution for efficient data‐intensive processing by building an optimized prototype MapReduce framework for Microsoft's Azure cloud platform on the basis of TomusBlobs. Finally, we specifically address the limitations of state‐of‐the‐art MapReduce frameworks for reduce‐intensive workloads, by proposing MapIterativeReduce as an extension of the MapReduce model. We validate the aforementioned contributions through large‐scale experiments with synthetic benchmarks and with real‐world applications on the Azure commercial cloud by using resources distributed across multiple data centers; they demonstrate that our solutions bring substantial benefits to data‐intensive applications compared with approaches relying on state‐of‐the‐art cloud object storage. Copyright © 2013 John Wiley & Sons, Ltd.

The global deployment of cloud datacenters is enabling large scale scientific workflows to improve performance and deliver fast responses. This unprecedented geographical distribution of the computation is doubled by an increase in the scale of the data handled by such applications, bringing new challenges related to the efficient data management across sites. High throughput, low latencies or cost-related trade-offs are just a few concerns for both cloud providers and users when it comes to handling data across datacenters. Existing solutions are limited to cloud-provided storage, which offers low performance based on rigid cost schemes. In turn, workflow engines need to improvise substitutes, achieving performance at the cost of complex system configurations, maintenance overheads, reduced reliability and reusability. In this paper, we introduce OverFlow, a uniform data management system for scientific workflows running across geographically distributed sites, aiming to reap economic benefits from this geo-diversity. Our solution is environment-aware, as it monitors and models the global cloud infrastructure, offering high and predictable data handling performance for transfer cost and time, within and across sites. OverFlow proposes a set of pluggable services, grouped in a data scientist cloud kit. They provide the applications with the possibility to monitor the underlying infrastructure, to exploit smart data compression, deduplication and geo-replication, to evaluate data management costs, to set a tradeoff between money and time, and optimize the transfer strategy accordingly. The system was validated on the Microsoft Azure cloud across its 6 EU and US datacenters. The experiments were conducted on hundreds of nodes using synthetic benchmarks and real-life bio-informatics applications (A-Brain, BLAST). The results show that our system is able to model accurately the cloud performance and to leverage this for efficient data dissemination, being able to reduce the monetary costs and transfer time by up to three times.

International audience

Infrastructure clouds revolutionized the way in which we approach resource procurement by providing an easy way to lease compute and storage resources on short notice, for a short amount of time, and on a pay-as-you-go basis. This new opportunity, however, introduces new performance trade-offs. Making the right choices in leveraging different types of storage available in the cloud is particularly important for applications that depend on managing large amounts of data within and across clouds. An increasing number of such applications conform to a pattern in which data processing relies on streaming the data to a compute platform where a set of similar operations is repeatedly applied to independent chunks of data. This pattern is evident in virtual observatories such as the Ocean Observatory Initiative, in cases when new data is evaluated against existing features in geospatial computations or when experimental data is processed as a series of time events. In this paper, we propose two strategies for efficiently implementing such streaming in the cloud and evaluate them in the context of an ATLAS application processing experimental data. Our results show that choosing the right cloud configuration can improve overall application performance by as much as three times.

The global deployment of cloud datacenters is enabling large web services to deliver fast response to users worldwide. This unprecedented geographical distribution of the computation also brings new challenges related to the efficient data management across sites. High throughput, low latencies, cost-or energy-related trade-offs are just a few concerns for both cloud providers and users when it comes to handling data across datacenters. Existing cloud data management solutions are limited to cloud-provided storage, which offers low performance based on rigid cost schemas. In this paper, we are proposing a dedicated cloud data transfer service that supports large-scale data dissemination across geographically distributed sites, advocating for a Transfer as a Service (TaaS) paradigm. The system aggregates the available bandwidth by enabling multiroute transfers across cloud sites. For users of multi-site or federated clouds, our proposal is able to decrease the variability of transfers and increase the throughput up to three times compared to baseline user options, while benefiting from the well-known high availability of cloud-provided services. For cloud providers, such a service can decrease the energy consumption within a datacenter down to half compared to user-based transfers.

Scientific workflows typically communicate data between tasks using files. Currently, on public clouds, this is achieved by using the cloud storage services, which are unable to exploit the workflow semantics and are subject to low throughput and high latencies. To overcome these limitations, we propose an alternative leveraging data locality through direct file transfers between the compute nodes. We rely on the observation that workflows generate a set of common data access patterns that our solution exploits in conjunction with context information to self-adapt, choose the most adequate transfer protocol and expose the data layout within the virtual machines to the workflow engines. This file management system was integrated within the Microsoft Generic Worker workflow engine and was validated using synthetic benchmarks and a real-life application on the Azure cloud. The results show it can bring significant performance gains: up to 5x file transfer speedup compared to solutions based on standard cloud storage and over 25% application timespan reduction compared to Hadoop on Azure.

The emergence of cloud computing brought the opportunity to use large-scale compute infrastructures for a broad spectrum of applications and users. As the cloud paradigm gets attractive for the " elasticity'' in resource usage and associated costs (the users only pay for resources actually used), cloud applications still suffer from the high latencies and low performance of cloud storage services. Enabling high-throughput massive data processing on cloud data becomes a critical issue, as it impacts the overall application performance. In this paper we address the above challenge at the level of the cloud storage. We introduce a concurrency-optimized data storage system which federates the virtual disks associated to VMs. We demonstrate the performance of our solution for efficient data-intensive processing on commercial clouds by building an optimized prototype MapReduce framework for Azure that leverages the benefits of our storage solution. We perform extensive synthetic benchmarks as well as experiments with real-world applications: they demonstrate that our solution brings substantial benefits to data intensive applications compared to approaches relying on state-of-the-art cloud object storage.

It has been shown that FPGAs could outperform high-end microprocessors on floating-point computations thanks to massive parallelism. However, most previous studies re-implement in the FPGA the operators present in a processor. This conservative approach is relatively straightforward, but it doesn't exploit the greater flexibility of the FPGA. We survey the many ways in which the FPGA implementation of a given floating-point computation can be not only faster, but also more accurate than its microprocessor counterpart. Techniques studied here include custom precision, mixing and matching fixed- and floating-point, specific accumulator design, dedicated architectures for coarser operators implemented as software in processors (such as elementary functions or Euclidean norms), operator specialization such as constant multiplication, and others. The FloPoCo project (http://www.ens-lyon.fr/LIP/Arenaire/Ware/FloPoCo/) aims at providing such non-standard operators. As a conclusion, current FPGA fabrics could be enhanced to improve floating-point performance. However, these enhancements should not take the form of hard FPU blocks as others have suggested. Instead, what is needed is smaller building blocks more generally useful to the implementation of floating-point operators, such as cascadable barrel shifters and leading zero counters

This article studies two common situations where the flexibility of FPGAs allows one to design application-specific floating-point operators which are more efficient and more accurate than those offered by processors and GPUs. First, for applications involving the addition of a large number of floating-point values, an ad-hoc accumulator is proposed. By tailoring its parameters to the numerical requirements of the application, it can be made arbitrarily accurate, at an area cost comparable to that of a standard floating-point adder, and at a higher frequency. The second example is the sum-of-product operation, which is the building block of matrix computations. A novel architecture is proposed that feeds the previous accumulator out of a floating-point multiplier whose rounding logic has been removed, again improving the area/accuracy tradeoff. These architectures are implemented within the FloPoCo generator, freely available under the LGPL.