|June 6, 2013
(Huxley 218, 2-3pm)
|Reaping the Benefits of Partial SDN Deployment in Enterprise Networks|
|May 24, 2013
(Huxley 218, 2-3pm)
|Osnat (Ossi) Mokryn
Tel Aviv Jaffa College
|The Power of Prediction: Cloud Bandwidth and Cost Reduction
|May 23, 2013
(William Penney 212, 2-3pm)
Royal Holloway, University of London
|Consistency and Complexity Tradeoffs for Highly-Available Multi-Cloud Store
|May 16, 2013
Imperial College London
|Discovering Service Dependencies in Mobile Ad Hoc Networks
|May 9, 2013||Abul Alim
|FlowOS: A programmable platform for writing middlebox software
|May 2, 2013
(Huxley 311, 2-3pm)
Microsoft Research Cambridge
|From High-level Haskel to Efficient Low-level Code
|April 25, 2013
INRIA & LIP6
|Swiftcloud: conflict-free objects at large scale
|April 22, 2013
University of Sydney
|Moana: A Case for a Global Information Network
|April 12, 2013||Florin Dinu
|Understanding and Improving the Efficiency of Failure Resilience for Big Data Frameworks
|April 11, 2013
|On the Naturalness of Software|
|April 10, 2013||Efthymia Tsamoura
Aristotle University of Thessaloniki
|Optimizing Queries in Modern Infrastructure: Challenges, Solutions and Open Problems|
|January 17, 2013||Alexandros Koliousis
Imperial College London
|Complex event processing for home network management
December 6, 2012
November 15, 2012
Raul Castro Fernandez
October 25, 2012
Mahesh K. Marina
October 11, 2012
October 4, 2012
September 13, 2012
Title: Reaping the Benefits of Partial SDN Deployment in Enterprise Networks
Speaker: Marco Canini (TU Berlin/T-Labs)
When:Thursday, 6/6/2013, time 2-3pm, room 218
Abstract: The operational challenges posed in enterprise networks, present an appealing opportunity for the software-defined orchestration of the network (SDN). However, the primary challenge to realizing solutions built on SDN in the enterprise is the deployment problem. Unlike in the datacenter, network upgrades in the enterprise start with the existing deployment and are particularly budget and resource-constrained. In this work, we investigate the prospect for partial Software Defined Network deployment. We present Panopticon, an architecture and methodology for planning and operating networks that combine legacy and upgraded SDN switches. Panopticon exposes an abstraction of a fully-deployed SDN in a partially upgraded legacy network, where the SDN benefits extend potentially over the entire network. We evaluate the feasibility of our approach through simulation on real enterprise campus network topologies entailing over 1500 switches and routers. Our results suggest that with only a handful of upgraded switches, it becomes possible to operate most of an enterprise network as a single SDN while meeting key resource constraints.
Short Bio: Marco is a senior research scientist at T-Labs, a joint institute of TU Berlin and Telekom Innovation Laboratories. Marco obtained his Ph.D. degree in Computer Science and Engineering from the University of Genoa in 2009 after spending the last year as a visiting student at the University of Cambridge, Computer Laboratory. He holds a laurea degree with honors in Computer Science and Engineering from the University of Genoa. He also held positions at Intel Research and Google, and he was a postdoctoral researcher at EPFL from 2009 to 2012.
Title: The Power of Prediction: Cloud Bandwidth and Cost Reduction
Speaker: Osnat (Ossi) Mokryn (Tel Aviv Jaffa College, Israel)
When: Friday, 24/05/2013, time 2-3pm, room 218
Abstract: In this paper we present PACK (Predictive ACKs), a novel end-to-end Traffic Redundancy Elimination (TRE) system, designed for cloud computing customers. Cloud-based TRE needs to apply a judicious use of cloud resources so that the bandwidth cost reduction combined with the additional cost of TRE computation and storage would be optimized. PACK's main advantage is its capability of offloading the cloud-server TRE effort to end-clients, thus minimizing the processing costs induced by the TRE algorithm. Unlike previous solutions, PACK does not require the server to continuously maintain client's status. This makes PACK very suitable for pervasive computation environments that combine client mobility and server migration to maintain cloud elasticity. PACK is based on a novel TRE technique, which allows the client to use newly received chunks to identify previously received chuck chains, which in turn can be used as reliable predictors to future transmitted chunks. We present a fully functional PACK implementation, transparent to all TCP-based applications and network devices. Finally, we analyze PACK benefits for cloud users, using traffic traces for various sources.
Joint work with Eyal Zohar and Israel Cidon. The work was presented at SigComm'11, and extended to a ToN paper 2013.
Short Bio: Osnat (Ossi) Mokryn received her B.Sc. in Computer Science and M.Sc. in Electrical Engineering from the Technion, Israel Institute of Technology, Haifa, Israel, in 1993 and 1998, respectively. She received her Ph.D in Computer Science from the Hebrew University, Jerusalem, Israel, in 2004. Ossi has worked at leading companies such as Intel and IBM research labs in Haifa, Israel. She was a post doctorate fellow in the Electrical Engineering Departments at Tel-Aviv University, Israel and also at the Technion, Haifa, Israel. Starting 2008 she is a faculty member at the School of Computer Science in Tel Aviv Jaffa College, Israel. Her recent research focuses on cloud content delivery and federated networks, data mining and recommenders in information systems and social networks.
Title: Consistency and Complexity Tradeoffs for Highly-Available Multi-Cloud Store
Speaker: Gregory Chockler (Royal Holloway, University of London)
When: Thursday, 23/05/2013, time 2-3pm, William Penney 212
Abstract: Cloud-based storage services have established themselves as a paradigm of choice for supporting bulk storage needs of modern networked services and applications. Although individual storage service providers can be trusted to do their best to reliably store the user data, exclusive reliance on any single provider or storage service leaves the users inherently at risk of being locked out of their data due to outages, connectivity problems, and unforeseen alterations of the service contracts. An emerging multi-cloud storage paradigm addresses these concerns by replicating data across multiple cloud storage services, potentially operated by distinct providers. In this paper, we study the impact of the storage interfaces and consistency semantics exposed by individual clouds on the complexity of the reliable multi-cloud storage implementation. Our results establish several inherent space and time tradeoffs associated with emulating reliable objects over a collection of unreliable storage services with varied interfaces and consistency guarantees.
Short Bio: Dr. Gregory Chockler is a Reader (Associate Professor) in the Department of Computer Science at the Royal Holloway, University of London (RHUL). He received the PhD degree in Computer Science from the Hebrew University of Jerusalem in 2003. He was a researcher with the Storage Systems Research group at IBM Research -- Haifa during 2002-2003. In 2003, he joined the Theory of Distributed Systems group in the MIT Computer Science and Artificial Intelligence Lab as a postdoctoral associate. He then returned to IBM Research -- Haifa as a Research Staff Member in 2005. He joined the CS faculty at RHUL in October 2012.
Dr. Chockler's research interests span all areas of distributed systems, and in particular, cloud computing, fault-tolerance, peer-to-peer computing, and middleware technologies. He regularly publishes and serves on conference organizing committees in these fields. He is an author or co-author of 10 patents, and more than 60 technical papers. His work at IBM was recognized through various corporate awards including 2010 and 2012 Outstanding Technical Accomplishment, 2011 Excellence and Eminence, and 2012 Scientific Accomplishment. He presently serves on the Editorial Board of Information Processing Letters, and was a guest editor of the special issue on Cloud Computing for Journal of Parallel and Distributed Computing (JPDC). He is a co-founder of the ACM/SIGOPS Workshop on Large-Scale Distributed Systems and Middleware (LADIS).
Title: Discovering Service Dependencies in Mobile Ad Hoc Networks
Speaker: Petr Novotny (Imperial College London)
When: Thursday, 16/05/2013, time 2-3pm, room W301
Abstract: The combination of service-oriented applications, with their run-times service bindings, and mobile ad hoc networks, with their transient communication topologies, brings a new level of complex dynamism to the structure and behavior of software systems. This complexity challenges our ability to understand the dependence relationships among system components when performing analyses such as fault localization and impact analysis. Current methods of dynamic dependence discovery, developed for use in fixed networks, assume that dependencies change slowly. Moreover, they require relatively long monitoring periods as well as substantial memory and communication resources, which are impractical in the mobile ad hoc network environment. We describe a new method, designed specifically for this environment, that allows the engineer to trade accuracy against cost, yielding dynamic snapshots of dependence relationships. We evaluate our method in terms of the accuracy of the discovered dependencies.
Title: FlowOS: A programmable platform for writing middlebox software
Speaker: Abdul Alim (Lancaster University)
When: Thursday, 09/05/2013, time 2-3pm, room TBA
Abstract: Middleboxes are ubiquitous in the Internet and are playing important role in today's networks ranging from home networks to ISPs to data centers. The use of distributed and web applications are increasing as a consequence of the popularity of cloud computing, which in tern demands more middlebox applications. However, existing middleboxes are closed hardware devices and difficult to program and upgrade. On the other hand, general purpose operating systems provide only low-level packet processing interfaces such as netfilter and pcap for writing middlebox applications. Many middlebox functions (e.g., application firewalls and application gateways, WAN optimizers, etc.) process application data instead of low-level IP packets.
In this talk, I would present a new programmable middlebox platform called FlowOS that provides a high-level interface for writing middlebox software. It hides the complexities of low-level packet processing, process synchronization, and inter-process communication. The preliminary tests show that FlowOS itself does not add any significant overhead to flow processing.
Short Bio: Abdul Alim has been working as a research associate at Lancaster University after receiving his PhD in Computer Science from the University of Cambridge in 2011. He graduated from the University of Rajshahi, Bangladesh in 2001 with an honors degree in Computer Science and worked there as a lecturer for four years. He received his MSc degree in Computer Science from the University of Northern British Columbia, Prince George, BC, Canada in 2007. His research interests fall in the area of distributed and networked systems in particular, but not limited to the design and analysis of routing protocols, network security, software defined networks, wireless ad hoc and sensor networking.
Title: From High-level Haskel to Efficient Low-level Code
Speaker: Geoffrey Mainland (MSRC)
When: Thursday, 02/05/2013, time 2-3pm, room 311
Abstract: Although high-level declarative languages like Haskell provide powerful abstraction facilities, conventional wisdom holds that such languages are inherently inefficient. I will show that we need not sacrifice abstraction for performance by demonstrating two systems that compile high-level Haskell to efficient low-level code. The first uses recent compiler and library advances to enable transforming numerical Haskell programs into fast machine code that exploits SSE and AVX instructions. The second compiles a subset of Haskell to run on GPUs.
Short Bio: Geoffrey Mainland obtained an A.B. in Physics and, in 2011, a Ph.D. in computer science from Harvard University under the supervision of Gregg Morrisett and Matt Welsh. He is presently a post doc with the Programming Principles and Tools group at Microsoft Research Cambridge. His research focuses on programming language and runtime support for non-general purpose computation.
Title: Swiftcloud: conflict-free objects at large scale
Speaker: Marc Shapiro (INRIA & LIP6)
When: Thursday, 25/04/2013, time 2-3pm, room 218
Abstract: Eventual Consistency (EC) is essential for availability and scalability of geo-replication, but existing EC approaches are ad-hoc and error-prone. Our Conflict-Free Replicated Data Types (CRDTs) are based on theoretically sound and simple sufficient conditions. We consider the design of a CRDT to ensure that it has intuitive semantics and that it uses memory efficiently. We also describe our Swiftcloud platform, designed to simplify the implementation of applications sharing objects at extreme scale, very close to clients at the network edge. Swiftcloud extends the CRDT guarantees (with transactional consistency, causal consistency, and session guarantees) without violating conflict-freedom. We describe some proof-of-concept applications (social networks, bookstore, file system) and report on their performance, which is greatly improved over the classical synchronisation-based approach.
Joint work with: Nuno Preguica (U. Nova de Lisboa), Marek Zawirski (INRIA & UPMC), Annette Bieniusa (INRIA & LIP6) and Carlos Baquero (U. Minho).
Short Bio: Marc Shapiro is a researcher in distributed and concurrent systems at INRIA in Paris (France). His research topics include large-scale distributed computing systems, focusing on data replication and consistency, especially in the wide area and in disconnected operation. After his PhD at LAAS (Toulouse), Marc Shapiro did his research at MIT (Cambridge, USA), CMIRH (Paris, France), INRIA (Rocquencourt, France), Cornell University (Ithaca, USA), Sun Microsystems (Chelmsford, USA), and Microsoft Research (Cambridge, UK). He is currently a senior research researcher for INRIA in the Regal group (INRIA-LIP6).
Home page: http://lip6.fr/Marc.Shapiro/
Title: Moana: A Case for a Global Information Network
Speaker: Yan Shvartzshnaider (University of Sydney and NICTA)
When: Monday, 22/04/2013, time 11-12pm, room 217
Abstract: There is an ever increasing gap between how the "common man" would describe the Internet and how a network researcher would. This is not suprising, as the "packet with a label indentifying the receiver" model seems to be an ill fill abstraction for popular services provided by companies such as Google, Facebook, YouTube, or Twitter. These services are primarily about information dissemination and a suitable network service abstraction should really do more than deliver individual packets with minimal latency.
We argue that we should adopt an information-centric system model instead of the current service oriented one. Rather than dealing with information exchange between specific services on a case by case basis, we should consider a global information storage and dissemination network on which individual applications and services are built.
In this talk I will discuss our vision for the future architecture and give a brief overview on our progress so far.
Short Bio: Yan Shvartzshnaider is a PhD Candidate in the school of Engineering and Information Technologies in the University of Sydney and a graduate researcher in National ICT Australia (NICTA), Networks Group in ATP lab. His research focus is on large scale networking systems, in particular, developing a fully distributed information-centric networking service abstraction.
Title: Understanding and Improving the Efficiency of Failure Resiliency for Big Data Frameworks
Speaker: Florin Dinu (Rice University)
When: Friday, 12/04/2013, time 2-3pm, room 217
Abstract: Big data processing frameworks (MapReduce, Hadoop, Dryad) are hugely popular today. A strong selling point is their ability to provide failure resilience guarantees. They can run computations to completion despite occasional failures in the system. However, an overlooked point has been the efficiency of the failure resilience provided. The vision of this work is that big data frameworks should not only finish computations under failures but minimize that impact of the failures on the computation time.
This part of the talk presents the first in-depth analysis of the efficiency of the failure resilience provided by the popular Hadoop framework at the level of a single job. The results show that compute node failures can lead to variable and unpredictable job running times. The causes behind these results are detailed in the talk. The second part of the talk focuses on providing failure resilience at the level of multi-job computations. It presents the design, implementation and evaluation of RCMP, a MapReduce system based on the fundamental insight that using replication as the main failure resilience strategy oftentimes leads to significant and unnecessary increases in computation running time. In contrast, RCMP is designed to use job re-computation as a first-order failure resilience strategy. RCMP enables re-computations that perform the minimum amount of work and also maximizes the efficiency of the re-computation work that still needs to be performed.
Short Bio: Florin Dinu is a final year graduate student in the Systems Group at Rice University, Houston, TX. He is advised by Prof. T.S. Eugene Ng. Before joining Rice in 2007, he received a B.A. in Computer Science from Politehnica University Bucharest in 2006 and then worked as a junior researcher at the Fokus Fraunhofer Institute in Berlin, Germany. His Ph.D dissertation focuses on the efficiency of failure resilience in big data processing frameworks. He has also done work on the benefits of centralized network control, congestion inference and improving data transfers for big data computations.
Title: On the Naturalness of Software
Speaker: Earl Barr (UCL)
When: Thursday, 11/04/2013, time 2-3pm, room 218
Abstract: Natural languages like English are rich, complex, and powerful, especially in the hands of masters like Shakespeare and Avvaiyar. Most human utterances, however, are far simpler, much more repetitive and predictable, due to cognitive limitations and the exigencies of daily life. In fact, modern statistical methods can very usefully model these utterances and have enjoyed phenomenal success when applied to speech recognition, natural language translation, question-answering, and text mining and comprehension.
We begin with the conjecture that most software is also natural, in the sense that it is created by humans at work, with all the attendant constraints and limitations --- and thus, like natural language, is also likely to be repetitive and predictable. We then ask whether statistical language modes can a) usefully model code and b) be leveraged to aid software engineers. Using the widely adopted N-gram model, we present empirical evidence supportive of positive answers to both questions. We show that code is very repetitive; in fact, even more so than natural languages. As an example use of the model, we developed a simple code completion engine for Java that, despite its simplicity, already improves Eclipse's built-in completion capability.
Short Bio: Earl Barr is a lecturer at the University College London. He received his M.S. (1999) and Ph.D. (2009) degrees, both in Computer Science, at the University of California at Davis. He was awarded the highly competitive I3P Fellowship from the Department of Homeland Security in 2010 and serves as a co-PI on three NSF grants and an Air Force DURIP grant. Dr. Barr's research interests include testing and analysis, empirical software engineering, computer security, and distributed systems. His recent work focuses on testing and analysis of numerical software, automated debugging, defect analysis and prediction, and code obfuscation.
Title: Optimizing Queries in Modern Infrastructure: Challenges, Solutions and Open Problems
Speaker: Efthymia Tsamoura (Aristotle University of Thessaloniki, AUTH, Greece)
When: Wednesday, 10/04/2013, time 2-3pm, room 217
Abstract: The nature of modern applications has rendered necessary the development of distributed environments, such as the grid and the cloud computing infrastructure. Despite the fact that query optimization/processing has been already studied since mid 1980s, the characteristics of modern infrastructure gave rise to novel problems that have not been addressed in the past. A number of these are presented in this talk, as well as, algorithms and techniques to tackle them developed during my PhD research. First, problems coming from the domain of distributed pipelined service ordering are discussed, while, next, a problem that is related to adaptive query processing is presented. Finally, two novel problems are introduced motivated by the fact that the resources in modern infrastructure are leased; both of them deal with extending state-of-the-art distributed or centralized platforms, the Mariposa and DRAGOON, respectively, on a cloud federation.
Short Bio: Efthymia Tsamoura received the Diploma degree in Informatics with honors in 2007 from Aristotle University of Thessaloniki (AUTH), Greece. Until 2007 she was a research assistant in the Artificial Intelligence and Information Analysis lab of AUTH, while from 2007 to 2008 she was a research assistant in the Information Technologies Institute, Center of Research and Technology Hellas. Currently she is pursuing a PhD in distributed, adaptive and multi-objective query optimization, Department of Informatics AUTH, under professor Yannis Manolopoulos. She is going to defend her thesis at May 2013. Efthymia Tsamoura received several national awards during both her undergraduate and postgraduate studies. Her main research interests lie in the areas of query optimization, cloud computing and statistical data management.
Title: Complex event processing for home network management
Speaker: Alexandros Koliousis (Imperial College London)
When: Thursday, 17/01/2012, time 2-3pm, room 217
Abstract: This talk describes a unification of publish/subscribe and stream database concepts for complex event processing. The unified system, originally designed as the centerpiece of a purpose-built home router, has enabled a wealth of user applications for controlling and understanding home networks. The proposed unity is expressed in a concise, yet powerful, query programming language-namely, the Glasgow Automaton Programming Language-that, in addition to home networking, it has been successfully applied to a diverse set of complex event detection scenarios.
Short Bio: Alexandros is a Research Associate at the Department of Computing, working on the HARNESS project (www.harness-project.eu). He is a member of the Experimental Software Systems and Large-Scale Distributed Systems group. Prior to joining Imperial College, he was a Research Assistant at the School of Computing Science, University of Glasgow-from where he also obtained his MSc and PhD degrees, under the supervision of Prof. Joe Sventek-working on the Homework project (www.homenetworks.ac.uk).
Title: Network Intrusion Detection and XML Processing
Speaker: Peter Ogden (Imperial College London)
When: Thursday, 6/12/2012, time 2-3pm, room 344B
Abstract: Network intrusion detection relies on looking for patterns in large numbers of concurrent sessions. Currently this involves reconstructing all of the flows on a link and using a large collection of regular expressions to hunt for malicious content. In this talk I will look at how processing data out of order can be used to reduce the amount of memory required for reconstructing flows and show that the same techniques can be applied to the processing of large XML data files in parallel.
Short Bio: Peter is a first year PhD working in the Circuits and Systems Group in EEE and LSDS in DoC. After completing his masters he spent 15 months working for BAE Systems Detica looking at high speed analysis of network traffic.
Title: Integrating Scale Out and Fault Tolerance in Stream Processing using Operator State Management
Speaker: Raul Castro Fernandez (Imperial College London)
When: Thursday, 15/11/2012, time 2-3pm, room 344B
Abstract: As users of big data applications expect fresh results, we witness a new breed of stream processing system (SPS) that are designed to scale to large numbers of cloud-hosted machines. Such systems face new challenges: (i) to benefit from the pay-as-you-go model of cloud computing, they much scale out on demand, acquiring additional virtual machines (VMs) and parallelising operators when the workload increases; (ii) failures are common with deployments on hundreds of VMs---systems must be fault-tolerant with fast recovery times, yet low per-machine overheads. An open question is how to achieve these two goals when stream queries include stateful operators, which be scaled out and recovered without affecting query results.
Our key idea is to expose internal operator state explicitly to the SPS through a set of state management primitives. Based on them, we describe a new integrated approach for dynamic scale out and recovery of stateful operators. Externalised operator state is checkpointed periodically by the SPS and backed up to upstreams VMs. The SPS then identifies individual operator bottlenecks and automaticaly scales them out by allocating new VMs and partitioning the checkpointed state from upstream VMs. At any point, failed operators are recovered by restoring checkpointed state on a new VM and replaying unprocessed tuples, not yet reflected in the state. We evaluate this approach with the Linear Road Benchmark on the Amazon EC2 cloud platform and show that it can scale automatically to a load factor of L=350 with 50 VMs, while recovering quickly from failure.
Title: Mobile Phone Sensing Broadly Defined
Speaker: Mahesh K. Marina (University of Edinburgh)
When: Thursday, 25/10/2012, time 2-3pm, room 344B
Abstract: The smartphone has emerged as an attractive sensing platform in recent years. This can be attributed to several factors, including rapid adoption of smartphones and their ubiquitous use, several built-in sensors they have (accelerometer, gyroscope, etc.), and ease of application deployment. Nevertheless, the scope of phone based sensing applications is apparently limited by the available set of sensors on the phone. It is in this context that taking a broader perspective on mobile phone sensing and viewing various wireless communication interfaces that are already present (or will be soon) on phones also as sensors can help overcome the above limitation and give rise to newer phone based sensing applications. To support this assertion, I will present a few case studies on environmental sensing; location sensing and network monitoring that rely on wireless communication interfaces on phones.
Short Bio: Mahesh K. Marina is a Lecturer in the School of Informatics at the University of Edinburgh. He received his Ph.D. in Computer Science from the State University of New York at Stony Brook and subsequently had a two-year postdoctoral stint at UCLA before joining Edinburgh. His current research spans rural/remote wireless access, mobile phone sensing, wireless network management and measurement.
Title: CloudFilter: Practical Control of Sensitive Data Propagation to the Cloud
Speaker: Ioannis Papagiannis (Imperial College London)
When: Thursday, 11/10/2012, time 2-3pm, room 344B
Abstract: A major obstacle for the adoption of cloud services in enterprises is the potential loss of control over sensitive data. Companies often have to safeguard a subset of their data because it is crucial to their business or they are required to do so by law. In contrast, cloud service providers handle enterprise data without providing guarantees and may put confidentiality at risk. In order to maintain control over their sensitive data, companies typically block all access to a wide range of cloud services at the network level. Such restrictions significantly reduce employee productivity while offering limited practical protection in the presence of malicious employees.
In this talk, we suggest a practical mechanism to ensure that an enterprise maintains control of its sensitive data while employees are allowed to use cloud services. We observe that most cloud services use HTTP as a transport protocol. Since HTTP offers well-defined methods to transfer files, inspecting HTTP messages allows the propagation of data between the enterprise and cloud services to be monitored independently of the implementation of specific cloud services. Our system, CloudFilter, intercepts file tranfers to cloud services, performs logging and enforces data propagation policies. CloudFilter controls where files propagate after they have been uploaded to the cloud and ensures that only authorised users may gain access. We show CloudFilter can be applied to control data propagation to Dropbox and GSS, describing the realistic data propagation policies that it can enforce.
This is a practice talk for a presentation at the ACM Cloud Computing Security Workshop (CCSW12).
Title: Fault Localization in MANET-Hosted Service-Based Systems
Speaker: Petr Novotny (Imperial College London)
When: Thursday, 04/10/2012, time 2-3pm, room 343
Abstract: Fault localization in general refers to a technique for identifying the likely root causes of failures observed in systems formed from components. Fault localization in systems deployed on mobile ad hoc networks (MANETs) is a particularly challenging task because those systems are subject to a wider variety and higher incidence of faults than those deployed in fixed networks, the resources available to track fault symptoms are severely limited, and many sources of the sources of faults in MANETs are by their nature transient. We present a method for localizing the faults occurring in service-based systems hosted on MANETs. The method is based on the use of dependence data that are discovered dynamically through decentralized observations of service interactions. We employ both Bayesian and timing-based reasoning techniques to analyze the data in the context of a specific fault propagation model, deriving a ranked list of candidate fault locations. We present the results of an extensive set of experiments exploring a wide range of operational conditions to evaluate the accuracy of our method.
This is a practice talk for a presentation at the 31st IEEE International Symposium on Reliable Distributed Systems.
Title: Overload Management in Data Stream Processing Systems with Latency Guarantees
Speaker: Evangelia Kalyvianaki (Imperial College London)
When: Thursday, 13/09/2012, time 2-3pm, room ***William Penny Building 212***
Abstract: Stream processing systems are becoming increasingly important to analyse real-time data generated by modern applications such as online social networks. Their main characteristic is to produce a continuous stream of fresh results as new data are being generated at real-time. Resource provisioning of stream processing systems is difficult due to time-varying workload data that induce unknown resource demands over time. Despite the development of scalable stream processing systems, which aim to provision for workload variations, there still exist cases such systems face transient resource shortages. During overload, there is a lack of resources to process all incoming data in real-time; data accumulate in memory and their processing latency grows uncontrollably compromising the freshness of stream processing results.
In this talk, I present a feedback control approach to design a nonlinear discrete-time controller that has no knowledge of the systems to be controlled or the workload for the data and is still able to control the average tuple end-to-end latency in a single-node stream processing system. Results, of our evaluation on a prototype stream processing system, show that our method controls the average tuple end-to-end latency despite the time-varying workload demands and increasing number of queries.
Short Bio: Eva Kalyvianaki is a Post-Doctoral Researcher at the Department of Computing, Imperial College London. Before this, she obtained her PhD from the Computer Laboratory, University of Cambridge. Her main research interests are in system, distributed systems, and autonomic computing.