Keynote talks

Keynote I:

Placing applications in large-scale distributed and heterogeneous environments

Prof Valeria Cardellini, University of Rome Tor-Vergata, Italy

Abstract: With the diffusion of new applications that require low latency and location awareness, there is a shift from running the applications in massive, centralized Cloud data centers with thousands of servers, to deploying and executing them in distributed Clouds comprised by many micro data centers equipped with heterogeneous computing and networking resources. In this context, an important problem is to decide where to place the different application components in such a way to achieve an efficient and network-aware usage of the underlying large-scale distributed infrastructure, while satisfying the Quality of Service (QoS) attributes of the application. Moreover, runtime adaptation capabilities are also required to cope with the highly changing environment in which the applications are executed. In this talk, I will provide some insights on the placement and deployment of applications with QoS requirements over a distributed infrastructure made of heterogeneous computing and networking resources. The considered applications range from distributed data stream processing to composite service-oriented applications, which are all executed in distributed cloud or edge computing environments. I will review some placement and runtime adaptation strategies that differ in their modeling assumptions and optimization goals, also considering their efficiency. I will also discuss some open research challenges which can provide interesting research hints to the MASCOTS community.

Biography: Valeria Cardellini is an associate professor in Computer Engineering at the University of Rome Tor-Vergata, Italy. Her main research interests focus on the design, engineering and performance evaluation of distributed systems and applications. In the last years she has been working on QoS-driven adaptation of service oriented applications, resource provisioning and pricing in Cloud systems, computation offloading in mobile Cloud computing, and QoS-aware placement of distributed data stream applications. She has published more than 80 refereed papers in international journals, book chapters, and conference proceedings and has edited two books. Three of these publications have received best paper awards. She contributes and has contributed actively in several international and national research projects, including the EU Cost Action ACROSS and the EoCoE Center of excellence. She received a PhD in Computer Science from the University of Rome Tor Vergata for work on scalable Web-server systems.

 


Keynote II:

Effectiveness: a new metric in performance evaluation

Dr G. Rubino, INRIA, France

Abstract: We present a performance metric called effectiveness that quantifies how well a stable queuing system in equilibrium works. Effectiveness captures a trade-off between three basic aspects of a system: its efficiency captured by its throughput, the cost in time to provide service measured by the mean response time, and the cost associated with the servers themselves. It is built on previous seminal work done by Kleinrock some years ago, and has some nice intuitive properties, in particular the fact that for basic queuing models such as the M/M/1 or the M/GI/1 queue, it is maximal when the mean number of customers in queue is one. Kleinrock calls this property “keep the pipe empty”. In the talk, the main characteristics of the effectiveness will be discussed, and, in particular, we show that the “keep the pipe empty” property also holds when the system is a Jackson product form network. We will also provide arguments to support the definition itself, showing that the metric allows the comparison of different systems and that it provides reasonable results. The connexion with Kleinrock’s work will be discussed, and we will end with some perspectives on the topic, including examples of open related questions.

Biography: Gerardo Rubino is a senior researcher at INRIA (the French National Institute for Research in Computer Science and Control) where he leads of the DIONYSOS group, working on the analysis and design of networking technologies and is a Board Member of the Media & Networks Cluster, Brittany, France. Among his past responsibilities, he has been Scientific Delegate for the Rennes unit of INRIA for five years, Research Director in Networking at the Telecom Bretagne engineering school for five years, Associate Editor of the Operations Research international journal “Naval Research Logistics” for nine years, former member of the Steering Board of the European Network of Excellence EuroFGI and its coordinator for industry collaborations during four years, and INRIA’s representative at the SISCom Brittany Research Cluster. He has also been the head of the International Partnership Office at INRIA Rennes. He is a member of IFIP WG 7.3 and has served on the Steering Committee of QEST (www.qest.org). He is interested in quantitative analysis of complex systems using probabilistic models, in networking and in other engineering areas. He presently works on performance and dependability analysis, and on perceptual quality assessment of audio and video applications and services built on top of the Internet. In particular, he is the author of the PSQA technology for automatic perceptual quality real-time evaluation (Pseudo-Subjective Quality Assessment). He also works on rare event analysis and is a member of the Steering Committee of RESIM, the only workshop dedicated to the topic, and co-authored the book “Rare Event Simulation Using Monte Carlo Methods” (Wiley, 2009). He is the author of more than 200 publications in applied mathematics and computer science, and recently (2014) co-authored the book “Markov Chains and Dependability Theory”, published by Cambridge University Press.

 


Keynote III:

Reducing the Wag in the Tail by Sharding and Replication

Prof Peter Harrison, Imperial College London, UK

Abstract: Access times can be reduced by splitting an object into fragments and distributing these onto devices that can be accessed in parallel. This has long been done in RAID storage systems by striping, for example. The processing paradigm for such systems is then essentially fork-join, for which the response time is the maximum of the fragment-response times, known to be a difficult problem in queueing networks. At there other extreme, replication of tasks - sending identical requests to multiple devices - and awaiting the first to complete improves response times by taking the minimum of task-response times. This is being used increasingly to reduce latency and the variability of response times in search engines, distributed systems and cloud computing. A disadvantage of fork-join is that it takes only one fragment to experience a severe delay, e.g. due to a wearing-out SSD cell, to cause a poor response time - in replication, such delays are relatively immaterial. Fork-join can be enhanced in systems where an object is split into N fragments, from any K

Biography: Peter Harrison is Professor of Mathematical Modelling in the Department of Computing at Imperial College London, where he has worked since 1976. He graduated at Christ's College Cambridge as a Wrangler in Mathematics in 1972 and went on to gain Distinction in Part III of the Mathematical Tripos in 1973, winning the Mayhew prize for Applied Mathematics. He has researched into stochastic performance modelling for some thirty-five years, visiting IBM Research Centres for two spells, written two books, published over 200 research papers, and chaired the Joint Sigmetrics/Performance conference at Imperial in 2012. The results of his research have been exploited extensively in industry, forming an integral part of Metron's Athene capacity planning tool. In the early 2000s he developed, and later mechanised, the RCAT methodology, from which new product-forms for Markovian networks have been found by several authors, and pioneered in the modelling of flash storage systems, incorporating dynamic workloads via Hidden Markov Models. Response time distributions have always been a favourite and are currently being investigated in sharded storage systems, task-replication and energy-saving, using both generating function and matrix analytic methods.