Our society is generating an increasing amount of data at an unprecedented scale, variety, and speed. This also applies to numerous research areas, such as genomics, high energy physics, and astronomy, for which large-scale data processing has become crucial. However, there is still a gap between the traditional scientific computing ecosystem and big data analytics tools and frameworks. On the one hand, high performance computing (HPC) programming models lack productivity, and do not provide means for processing large amounts of data in a simple manner. On the other hand, existing big data processing tools have performance issues in HPC environments, and are not general-purpose. In this paper, we propose and evaluate PyCOMPSs, a task-based programming model for Python, as an excellent solution for distributed big data processing in HPC infrastructures. Among other useful features, PyCOMPSs offers a highly productive general-purpose programming model, is infrastructure-agnostic, and provides transparent data management with support for distributed storage systems. We show how two machine learning algorithms (Cascade SVM and K-means) can be developed with PyCOMPSs, and evaluate PyCOMPSs’ productivity based on these algorithms. Additionally, we evaluate PyCOMPSs performance on an HPC cluster using up to 1,536 cores and 320 million input vectors. Our results show that PyCOMPSs achieves similar performance and scalability to MPI in HPC infrastructures, while providing a much more productive interface that allows the easy development of data analytics algorithms.

Performance anomaly detection is crucial for long running, large scale distributed systems. However, existing works focus on the detection of specific types of anomalies, rely on historical failure data, and cannot adapt to changes in system behavior at run time. In this work, we propose an adaptive framework for the detection and identification of complex anomalous behaviors, such as deadlocks and livelocks, in distributed systems without historical failure data. Our framework employs a two-step process involving two online SVM classifiers on periodically collected system metrics to identify at run time normal and anomalous behaviors such as deadlock, livelock, unwanted synchronization, and memory leaks. Our approach achieves over 0.70 F-score in detecting previously unseen anomalies and 0.78 F-score in identifying the type of known anomalies with a short delay after the anomalies appear, and with minimal expert intervention. Our experimental analysis uses system execution traces from our in-house distributed system with varied behaviors and a dataset by Yahoo!, and shows the benefits of our approach as well as future research challenges.

In recent years, machine learning has proven to be an extremely useful tool for extracting knowledge from data. This can be leveraged in numerous research areas, such as genomics, earth sciences, and astrophysics, to gain valuable insight. At the same time, Python has become one of the most popular programming languages among researchers due to its high productivity and rich ecosystem. Unfortunately, existing machine learning libraries for Python do not scale to large data sets, are hard to use by non-experts, and are difficult to set up in high performance computing clusters. These limitations have prevented scientists to exploit the full potential of machine learning in their research. In this paper, we present and evaluate dislib, a distributed machine learning library on top of PyCOMPSs programming model that addresses the issues of other existing libraries. In our evaluation, we show that dislib can be up to 9 times faster, and can process data sets up to 16 times larger than other popular distributed machine learning libraries, such as MLlib. In addition to this, we also show how dislib can be used to reduce the computation time of a real scientific application from 18 hours to 17 minutes.

Progress in science is deeply bound to the effective use of high-performance computing infrastructures and to the efficient extraction of knowledge from vast amounts of data. Such data comes from different sources that follow a cycle composed of pre-processing steps for data curation and preparation for subsequent computing steps, and later analysis and analytics steps applied to the results. However, scientific workflows are currently fragmented in multiple components, with different processes for computing and data management, and with gaps in the viewpoints of the user profiles involved. Our vision is that future workflow environments and tools for the development of scientific workflows should follow a holistic approach, where both data and computing are integrated in a single flow built on simple, high-level interfaces. The topics of research that we propose involve novel ways to express the workflows that integrate the different data and compute processes, dynamic runtimes to support the execution of the workflows in complex and heterogeneous computing infrastructures in an efficient way, both in terms of performance and energy. These infrastructures include highly distributed resources, from sensors and instruments, and devices in the edge, to High-Performance Computing and Cloud computing resources. This paper presents our vision to develop these workflow environments and also the steps we are currently following to achieve it.

The diagnosis, prediction, and understanding of unexpected behavior is crucial for long running, large scale distributed systems. However, existing works focus on the identification of faults in specific time moments preceded by significantly abnormal metric readings, or require a previous analysis of historical failure data. In this work, we propose an online behavior classification system to identify a wide range of undesired behaviors, which may appear even in healthy systems, and their evolution over time. We employ a two-step process involving two online classifiers on periodically collected system metrics to identify at runtime normal and anomalous behaviors such as deadlock, starvation and livelock, without any previous analysis of historical failure data. Our approach achieves over 80% accuracy in detecting unexpected behaviors and over 90% accuracy in identifying their type with a short delay after the anomalies appear, and with minimal expert intervention. Our experimental analysis uses system execution traces obtained from a Google cluster and from our in-house distributed system with varied behaviors, and shows the benefits of our approach as well as future research challenges.

Cloud Computing is a new computing paradigm where a large amount of computing capacity is offered on demand and only paying for what you use. Several Infrastructure Providers have adopted this approach offering resources which are easily managed by means of web-based APIs. However, if a user wants to use different providers, the resource management becomes tedious because providers define different API requiring a special implementation for interacting with each of them. In this paper, we present a methodology for making the provider interoperability easier. In this methodology, each provider's API is modeled by an ontology. Equivalences between these ontologies are modeled by rules, and messages used in a provider's API are converted in calls to another provider's API applying these rules. With our approach, users interact with Infrastructure Providers using their most familiar API and the translation to the other APIs is automatically done by the system.

Cloud computing is inherently service-oriented: cloud applications are delivered to consumers as services via the Internet. Therefore, these applications can potentially benefit from the Service-Oriented Architecture (SOA) principles: they can be programmed as added-value services composed by pre-existing ones, thus favouring code reuse. However, new programming models are required to simplify their development, along with systems that are capable of orchestrating the execution of the resulting SaaS in the Cloud. In that regard, this paper presents Service Super scalar (Servicess), an alternative to existing PaaS which provides a programming model and execution runtime to ease the development and execution of service-based applications in clouds. Servicess is a task-based model: the user is only required to select the tasks, which can be services or regular methods, to be spawned asynchronously. The application, a composite service, is programmed in a totally sequential way and no API call must be included in the code. The runtime is in charge of automatically orchestrating the execution of the tasks in the Cloud, as well as of elastically deploying new virtual resources depending on the load. After describing the main characteristics of the programming model and the runtime, we evaluate the productivity of Servicess and show how it offers a good trade-off between programmability and runtime performance.