Section: New Results

Performance Evaluation

Participants: Yann Busnel, Yves Mocquard, Bruno Sericola, Gerardo Rubino

Correlation estimation between distributed massive streams. The real time analysis of massive data streams is of utmost importance in data intensive applications that need to detect as fast as possible and as efficiently as possible (in terms of computation and memory space) any correlation between its inputs or any deviance from some expected nominal behavior. The IoT infrastructure can be used for monitoring any events or changes in structural conditions that can compromise safety and increase risk. It is thus a recurrent and crucial issue to determine whether huge data streams, received at monitored devices, are correlated or not as it may reveal the presence of attacks. In [14] we propose a metric, called Codeviation, that allows to evaluate the correlation between distributed massive streams. This metric is inspired from classical material in statistics and probability theory, and as such enables to understand how observed quantities change together, and in which proportion. We then propose to estimate the codeviation in the data stream model. In this model, functions are estimated on a huge sequence of data items, in an online fashion, and with a very small amount of memory with respect to both the size of the input stream and the domain from which data items are drawn. We then generalize our approach by presenting a new metric, the Sketch-$☆$ metric, which allows us to define a distance between updatable summaries of large data streams. An important feature of the Sketch-$☆$ metric is that, given a measure on the entire initial data streams, the Sketch-$☆$ metric preserves the axioms of the latter measure on the sketch. We also conducted extensive experiments on both synthetic traces and real data sets allowing us to validate the robustness and accuracy of our metrics.

Stream processing systems. Stream processing systems are today gaining momentum as tools to perform analytics on continuous data streams. Their ability to produce analysis results with sub-second latencies, coupled with their scalability, makes them the preferred choice for many big data companies.

A stream processing application is commonly modeled as a direct acyclic graph where data operators, represented by nodes, are interconnected by streams of tuples containing data to be analyzed, the directed edges (the arcs). Scalability is usually attained at the deployment phase where each data operator can be parallelized using multiple instances, each of which will handle a subset of the tuples conveyed by the operators’ ingoing stream. Balancing the load among the instances of a parallel operator is important as it yields to better resource utilization and thus larger throughputs and reduced tuple processing latencies.

Shuffle grouping is a technique used by stream processing frameworks to share input load among parallel instances of stateless operators. With shuffle grouping each tuple of a stream can be assigned to any available operator instance, independently from any previous assignment. A common approach to implement shuffle grouping is to adopt a Round-Robin policy, a simple solution that fares well as long as the tuple execution time is almost the same for all the tuples. However, such an assumption rarely holds in real cases where execution time strongly depends on tuple content. As a consequence, parallel stateless operators within stream processing applications may experience unpredictable unbalance that, in the end, causes undesirable increase in tuple completion times. In [61] we propose Online Shuffle Grouping (OSG), a novel approach to shuffle grouping aimed at reducing the overall tuple completion time. OSG estimates the execution time of each tuple, enabling a proactive and online scheduling of input load to the target operator instances. Sketches are used to efficiently store the otherwise large amount of information required to schedule incoming load. We provide a probabilistic analysis and illustrate, through both simulations and a running prototype, its impact on stream processing applications.

Grand Challenge. Since 2011, the ACM International Conference on Distributed Event-based Systems (DEBS) launched the Grand Challenge series to increase the focus on these systems as well as provide common benchmarks to evaluate and compare them. The ACM DEBS 2017 Grand Challenge focused on (soft) real-time anomaly detection in manufacturing equipment. To handle continuous monitoring, each machine is fitted with a vast array of sensors, either digital or analog. These sensors provide periodic measurements, which are sent to a monitoring base station. The latter receives then a large collection of observations. Analyzing in an efficient and accurate way, this very-high-rate – and potentially massive – stream of events is the core of the Grand Challenge. Although, the analysis of a massive amount of sensor reading requires an on-line analytics pipeline that deals with linked-data, clustering as well as a Markov model training and querying. The FlinkMan system [62] proposes a solution to the 2017 Grand Challenge, making use of a publicly available streaming engine and thus offering a generic solution that is not specially tailored for this or for another challenge. We offer an efficient solution that maximally utilizes available cores, balances the load among the cores, and avoids to the extent possible tasks such as garbage collection that are only indirectly related to the task at hand.

Health big data processing. Sharing and exploiting efficiently Health Big Data (HBD) lead to tackle great challenges: data protection and governance taking into account legal, ethical and deontological aspects which enables a trust, transparent and win-to-win relationship between researchers, citizen and data providers. Lack of interoperability: data are compartmentalized and are so syntactically and semantically heterogeneous. Variable data quality with a great impact on data management and statistical analysis. The objective of the INSHARE project [41] is to explore, through an experimental proof of concept, how recent technologies could overcome such issues. It aims at demonstrating the feasibility and the added value of an IT platform based on CDW, dedicated to collaborative HBD sharing for medical research.

The consortium includes 6 data providers: 2 academic hospitals, the SNIIRAM (the French national reimbursement database) and 3 national or regional registries. The platform is designed following a three steps approach: (1) to analyze use cases, needs and requirements, (2) to define data sharing governance and secure access to the platform, (3) to define the platform specifications. Three use cases (healthcare trajectory analysis, epidemiological registry enrichment, signal detection) were analyzed to design the platform corresponding to five studies and using eleven data sources. The governance was derived from the SCANNER model and adapted to data sharing. As a result, the platform architecture integrates the following tools and services: data repository and hosting, semantic integration services, data processing, aggregate computing, data quality and integrity monitoring, id linking, multi-source query builder, visualization and data export services, data governance, study management service and security including data watermarking.

Throughput prediction in cellular networks. Downlink data rates can vary significantly in cellular networks, with a potentially non-negligible effect on the user experience. Content providers address this problem by using different representations (e.g., picture resolution, video resolution and rate) of the same content and by switching among these based on measurements collected during the connection. If it were possible to know the achievable data rate before the connection establishment, content providers could choose the most appropriate representation from the very beginning. We have conducted a measurement campaign involving 60 users connected to a production network in France, to determine whether it is possible to predict the achievable data rate using measurements collected, before establishing the connection to the content provider, on the operator’s network and on the mobile node. We show that it is indeed possible to exploit these measurements to predict, with a reasonable accuracy, the achievable data rate [53].

Population protocol model. We consider in [50] a large system populated by $n$ anonymous nodes that communicate through asynchronous and pairwise interactions. The aim of these interactions is, for each node, to converge toward a global property of the system that depends on the initial state of the nodes. We focus on both the counting and proportion problems. We show that for any $\delta \in (0,1)$, the number of interactions needed per node to converge is $O(ln(n/\delta \left)\right)$ with probability at least $1-\delta$. We also prove that each node can determine, with any high probability, the proportion of nodes that initially started in a given state without knowing the number of nodes in the system. This work provides a precise analysis of the convergence bounds, and shows that using the 4-norm is very effective to derive useful bounds.

The context of [71] is the well studied dissemination of information in large scale distributed networks through pairwise interactions. This problem, originally called rumor mongering, and then rumor spreading has mainly been investigated in the synchronous model, which relies on the assumption that all the nodes of the network act in synchrony, that is, at each round of the protocol, each node is allowed to contact a random neighbor. In this paper, we drop this assumption under the argument that it is not realistic in large scale systems. We thus consider the asynchronous variant, where, at random times, nodes successively interact by pairs exchanging their information on the rumor. In a previous paper, we performed a study of the total number of interactions needed for all the nodes of the network to discover the rumor. While most of the existing results involve huge constants that do not allow us to compare different protocols, we provided a thorough analysis of the distribution of this total number of interactions together with its asymptotic behavior. In this paper we extend this discrete-time analysis by solving a conjecture proposed previously and we consider the continuous-time case, where a Poisson process is associated with each node to determine the instants at which interactions occur. The rumor spreading time is thus more realistic since it is the time needed for all the nodes of the network to discover the rumor. Once again, as most of the existing results involve huge constants, we provide a tight bound and equivalent of the complementary distribution of the rumor spreading time. We also give the exact asymptotic behavior of the complementary distribution of the rumor spreading time around its expected value when the number of nodes tends to infinity.

Transient analysis. Last, in two keynotes ([35] and [34]), we described part of our previous analytical results concerning the transient behavior of well-structured Markov processes, mainly on performance models (queueing systems), and we presented recent new results that extend those initial findings. The heart of the novelties lie on an extension of the concept of duality proposed by Anderson in  [73] that we call pseudo-dual. The dual of a stochastic process needs strong monotonicity conditions to exist. Our proposed pseudo-dual always exist, and is directly defined on a linear system of differential equations with constant coefficients, that can be, in particular, the system of Chapman-Kolmogorov equations corresponding to a Markov process, but not necessarily. This allows, for instance, to prove the validity of closed-forms expressions of the transient distribution of a Markov process in cases where the dual doesn't exist. The keynote [35] was presented to a public oriented toward differential equations and dynamical systems; [34] has a more modeling flavour. A paper is under preparation with the technical details.