ASF Winter School 2017

Failure detection in large and dynamic distributed systems

Nowadays, distributed systems are more and more large and versatile. Computing units can join, leave or move inside a global infrastructure. These features require the implementation of dynamic systems that can cope autonomously with changes in their structure. It therefore becomes necessary to define, develop, and validate distributed algorithms able to manage such dynamic at a large scale. Failure detection is a prerequisite to failure mitigation and a key component to build distributed algorithms requiring resilience. We introduce the problem of failure detection in asynchronous network where the transmission delay is not known. Then, we focus on new advances for scaling and taking into account the dynamic of the infrastructure. We present a new scalable perfect failure detection for typical HPC configurations and a new leader election algorithm for dynamic and mobile unknown networks.

Speaker's bio: Pierre Sens received his Ph. D. in Computer Science in 1994, and the “Habilitation à diriger des recherches” in 2000 from Paris 6 University (UPMC), France. Currently, he is a full Professor at Université Pierre et Marie Curie. His research interests include distributed systems and algorithms, large scale data storage, fault tolerance, and cloud computing. Since 2005, Pierre Sens is heading the Regal group which is a joint research team between LIP6 and Inria Paris. He was member of the Program Committee of of major conferences in the areas of distributed systems and parallism (ICDCS, IPDPS, OPODIS, ICPP, Europar,…) and serves as General chair of SBAC and EDCC. Overall, he has published over 120 papers in international journals and conferences.

Multicore chips: finally, shared memory distributed algorithms and simulations are practically useful

Shared memory algorithms have been extensively studied in the distributed computing community for decades now. With the advent of multicore chips, many concepts and algorithms from this line of research, such as lock-free or wait-free algorithms and simulations, shed a particular light on how to produce efficient code for multicore, and where do synchronisation bottlenecks reside.

Speaker's bio:

System techniques to mitigate the NUMA effect

Large multicores have a complex memory topology. They rely on a Non Uniform Memory Access (NUMA) architecture, which connects a set of memory banks, CPUs and caches with a high-speed network called "interconnect". Because of the distributed nature of this topology, an inadequate data placement overloads the interconnect, which drastically reduces memory access latency. Through the study of two system components, a garbage collector and an hypervisor, this talk presents techniques to mitigate this effect.

Speaker's bio: Gaël Thomas is professor at Telecom SudParis where he leads the HP2 group of the computer science department. His research aims at improving the performance, the design and the safety of the system stack. He received the PhD degree in 2005 and the "Habilitation à diriger les recherche" in 2012, both from UPMC Sorbonne Université. Before joining Telecom SudParis in 2014, he was associate professor at UPMC Sorbonne Université from 2006 to 2014.

Secure code execution in the Cloud

Cloud infrastructures has become the environment of choice for hosting networked services, executing time-consuming tasks, or storing massive amounts of data. Companies are however still reluctant to use Cloud services for sensitive tasks, as they could easily be compromised by technical vulnerabilities or malevolent insiders. Trusted execution environments (TEEs), such as Intel's SGX, provide a practical solution to this problem. They rely on dedicated hardware in the processor to protect selected code and data from disclosure or modification. TEEs have recently attracted much attention from the research community as they combine security and efficiency but still leave many challenging questions open for exploration. This presentation will give an overview of techniques for secure code execution, and in particular of Intel's SGX extensions found in its recent processors.

Speaker's bio: Pascal Felber received his M.Sc. and Ph.D. degrees in Computer Science from the Swiss Federal Institute of Technology. From 1998 to 2002, he has worked at Oracle Corporation and Bell-Labs (Lucent Technologies) in the USA. From 2002 to 2004, he has been an Assistant Professor at Institut EURECOM in France. Since October 2004, he is a Professor of Computer Science at the University of Neuchâtel, Switzerland, working in the field of dependable, concurrent, and distributed computing. He has published over 150 research papers in various journals and conferences

Tackling the video delivery challenge through system and network virtualization (using SDN/NFV)

It is foreseen that more than 82% of all Internet traffic will be video in 2020 and it will continue to increase. Nowadays, most of the video over the Internet is being distributed by Content Delivery Networks techniques, Over-The-Top of the Network Operators, making them loose all the control over this content. Continuing this way will not permit to achieve a sufficient Quality of Experience for consumers and will lack capabilities to scale. The emergence of virtualization techniques, in-system and in-network using Software Defined Networks (SDN) and Network Function Virtualization (NFV), provides interesting approaches to tackle the important challenge of video delivery, while at the same time bringing back the operators into the game and new actors as well (e.g., IOT providers, end-users) for the great benefit of the service and the whole value chain.

Speaker's bio: Daniel NEGRU (Ph.D) received his MSc. Degree in Computer Science from the University of Pierre and Marie Curie, Paris 6, in 2002. He worked in the centre of research of Motorola, Paris, specifying and developing a complete IPv6 stack from scratch, including Mobile IPv6 and multimedia components. Between 2003 and 2006, he worked at the CNRS-PRiSM laboratory, Paris area – French computer science research centre that does research in the underlying technologies for the next generation global information infrastructure – and participated in several national and European projects, such as ANR RIAM NMS, IST FP6 ATHENA, IST FP6 ENTHRONE, IST FP6 ENTHRONE2, IST FP6 IMOSAN. He received his PhD in 2006 in the field of Broadcast and Internet convergence solutions at the network and service levels. In 2007, he became Associate Professor at ENSEIRB-MATMECA School of Engineers / University of Bordeaux associated to the CNRS-LaBRI lab, specializing in multimedia and networking and his main fields of research resided in the domains of mobility and multimedia over heterogeneous networks, service accessibility and adaptation. He has next concentrated more efforts in media- aware networks, including content- & context-awareness and content adaptation, especially considering video streaming. He has coordinated the ICT FP7 ALICANTE project that tackles these fields (2010-2014). From then, he has expanded his area of competencies into network virtualization targeting NFV/SDN solutions for a virtual CDN and a virtual Home Gateway, while continuing on focusing on innovative video streaming solutions, such as a novel multiple description approach (MD-DASH). Towards these, he has participated to ICT FP7 T-NOVA and CHIST-ERA DISEDAN projects. He has published more than 50 papers, including high level ones in prestigious journals and conferences, such IEEE Communication Magazine, IEEE Multimedia Magazine, Elsevier, Globecom, ICC, FIA, etc. In 2013, he received his Habilitation à Diriger des Recherches (HDR) from the University of Bordeaux.

Is Crowdsourcing Yet Another Distributed System?

In crowdsourcing, tasks are completed by a network of resources, in this case humans, and a central entity, in this case a requester, is in charge of gathering and aggregating individual results. We are therefore tempted to think of crowdsourcing as another instance of a distributed system. However, humans play a central role in crowdsourcing. The crowd is typically volatile, its arrival and departure asynchronous, and its levels of attention and accuracy diverse. Human factors such as skills, expected wage and motivation, raise new

challenges in the distributed completion of tasks. In this talk, we first make a case for why crowdsourcing is yet another distributed system. We then describe how human factors affect task assignment and completion. We conclude with several open research questions.

This is joint work with my student Julien Pilourdault and with colleagues Senjuti Basu Roy from the New Jersey Institute of Technology and Dongwon Lee from Penn State University.

Speaker's bio: Sihem Amer-Yahia is DR1 CNRS at LIG in Grenoble where she leads the SLIDE team. Her interests are at the intersection of large-scale data management and data analytics. Before joining CNRS, she was Principal Scientist at QCRI, Senior Scientist at Yahoo! Research and Member of Technical Staff at at&t Labs. Sihem served on the SIGMOD Executive Board, the VLDB Endowment, and the EDBT Board. She is Editor-in-Chief of the VLDB Journal and is on the editorial boards of ACM TODS and the Information Systems Journal. Sihem is PC chair for VLDB 2018. Sihem received her Ph.D. in CS from Paris-Orsayand INRIA in 1999, and her Diplômed'ingénieur from INI, Algeria.