direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Inhalt des Dokuments

Talks, student talks, and other events of 2015

Tuesday, 15. December 2015

User driven traffic prioritization in WiFi networks through participatory networking, following an SDN and NFV approach
Speaker:
Carlos Mayer (TU Berlin master student)
Type:
Masterstudent final talk
Time:
15 December 2015
17:00–18:00
Place:
MAR 4.033
Abstract:
Networks in the customer’s premise must support a very diverse set of devices and require- ments. For example, some applications are bandwidth hungry and latency critical (e.g. live video calls) whereas other are not (e.g. Dropbox). Today, however, such networks are agnostic to different devices, applications and user preferences, but rather treat every flow equally. In this thesis, we envision a big potential to improve network performance and user satisfaction by explicitly taking into account heterogeneity and user’s preferences. Our main contribution is a network architecture that abstracts low-level network details and allows users to express their service preferences, through a participatory interface. Our architecture follows the SDN and NFV paradigm. Specifically, we leverage deep-packet-inspection to classify flows in order to program the data plane according to the user’s preferences. We provide a proof-of-concept implementation that successfully fulfills our requirements and report on an initial evaluation that shows promising results for long-lived flows.

Tuesday, 08. December 2015

ISmart mHTTP: Chunk Scheduler Design and Implementation for the Multi-source Multipath HTTP
Speaker:
Karl Fischer (TU Berlin master student)
Type:
Masterstudent final talk
Time:
08 December 2015
14:00–15:00
Place:
MAR 4.033
Abstract:
Multi-source Multipath HTTP (mHTTP) is a client oriented mechanism which enables users to simultaneously download partial chunks of the same content from different servers. The previous implementation of mHTTP, i. e., mHTTP Vanilla, has proven that it takes advantages of different types of network diversity, e. g., Interface diversity, data source diversity, and path diversity, for downloading a large file. However, how mHTTP performs on web pages, i. e., downloading a HTML page followed by multiple embedded web objects, is largely uncharted in the previous study. Furthermore, mHTTP Vanilla applies a fixed size to all partial blocks of the content during the transfer, which leads to performance issues when the quality of the links changes. Link performance changes should be handled by an intelligent algorithm, i. e., a scheduler that minimizes the request overhead, while still being adaptive to quality changes of the link. Additionally, a path manager which is part of the scheduler should select an optimal path, meaning the determination of a client interface and a server in order to achieve maximum overall throughput. The major purpose of this thesis is to develop and evaluate efficient scheduling algorithms for mHTTP. Key challenges to perform this study are: First, to develop Smart mHTTP, i. e., a prototype that allows multi-file (web page) downloads and a dynamic chunk size. Second, to design and implement different scheduling algorithms for Smart mHTTP. Finally, to evaluate these scheduling algorithms over controlled testbeds and in real-world experiments.

Tuesday, 24. November 2015

Predictable Performance for Map Reduce Jobs
Speaker:
Jan Henning (TU Berlin master student)
Type:
Masterstudent introductory talk
Time:
09 September 2015
14:00–15:00
Place:
MAR 4.033
Abstract:
Jan is currently extending Hadoop with the goal to create a proof of concept implementation of dynamic resource allocations in Hadoop. He intends to derive models to predict the (remaining) runtime of Hadoop jobs and enable the users to specify deadlines for their jobs. The cluster should tune resource allocations automatically to meet the requested deadlines.

Tuesday, 03. November 2015

How Hard Can It Be? Understanding the Complexity of Replica Aware Virtual Cluster Embeddings
Speaker:
Carlo Fürst
Type:
Project group meeting (Projektgruppentreffen, PGT)
Time:
03 November 2015
15:00 - 16:00
Place:
MAR 4.033
Abstract:

Virtualized datacenters offer great flexibilities in terms of resource allocation. In particular, by decoupling applications from the constraints of the underlying infrastructure, virtualization supports an optimized mapping of virtual machines as well as their interconnecting network to their physical counterparts: essentially a graph embedding problem. However, existing embedding algorithms such as Oktopus and Proteus often ignore a crucial dimension of the embedding problem, namely data locality: the input to a cloud application such as MapReduce is typically stored in a distributed, and sometimes redundant, file system. Since moving data is costly, an embedding algorithm should be data locality aware, and allocate computational resources close to the data; in case of redundant storage, the algorithm should also optimize the replica selection. This paper initiates the algorithmic study of data locality aware virtual cluster embeddings on datacenter topologies. We show that despite the multiple degrees of freedom in terms of embedding, replica selection and assignment, many problems can be solved efficiently. We also highlight the limitations of such optimizations, by presenting several NP-hardness proofs; interestingly, our hardness results also hold in uncapacitated networks of small diameter.

Tuesday, 27. October 2015

Investigating Rate Control in the Context of Replica Selection to reduce Tail Latencies
Speaker:
Tobias Brandt (TU Berlin master student)
Type:
Masterstudent final talk
Time:
27 October 2015
15:00–16:00
Place:
MAR 4.033
Abstract:
The recently proposed C3 system uses adaptive replica selection for reducing latency in the context of replicated and partitioned data-stores. In order to ensure that a set of clients do not saturate individual servers in the system, C3 uses distributed rate-control, wherein clients adapt their sending rates to servers in a decentralized fashion. The original C3 system used an adaptation of the TCP-CUBIC congestion control algorithm in order to perform distributed rate-control. The goal of this thesis is to investigate alternative rate adaptation mechanisms for the C3 system. In particular, we will focus on the i) latency based rate-adaptation mechanisms of the PARDA and PISCES systems (inspired by FAST-TCP), ii) the BIC mechanism used by the PULSAR system, as well as iii) a simple AIMD approach. The evaluation showed that after several adaptations PARDA/PISCES performed very well in specific scenarios but not in general. AIMD and BIC TCP did not introduce significant performance differences to C3's original cubic rate adaptation function.

Thursday, 22. October 2015

Substrate-Aware SDN Abstractions
Speaker:
Apoorv Shukla
Type:
Project group meeting (Projektgruppentreffen, PGT)
Time:
22 October 2015
15:00 - 16:00
Place:
MAR 4.033
Abstract:

Arguably one of the key benefits of today’s trend towards more softwarized and virtualized computer networks is the possibility to provide abstractions i.e. a logical network view which decouples the application from the constraints (and possibly confidential details) of the underlying physical infrastructure. While the benefits of such abstractions have been acknowledged and well-explored in the literature (e.g., for resource optimization or network virtualization), interestingly, the fundamental question of what is a good abstraction has not received much attention. In particular, existing systems often explicitly or implicitly rely on a “big switch” abstraction. While a big switch can be attractive for its simplicity, we, argue that it sometimes hides the crucial and necessary details required for fine-grained traffic engineering and enforcement of complex policies, as well as for providing complex service chains traversing a series of (physical or virtualized) middleboxes in a correct order. In this paper, we initiate a discussion on preserving the native semantics of the substrate, and providing abstractions which take into account essential characteristics such as connectivity, bandwidth, latency and the location of middleboxes. We propose and implement novel logical abstractions, which exploit the capabilities of an intermediate abstraction, which is as close to a big switch as possible but preserves the details of substrate-level complexity. The new approach enables predictable services over our logical abstractions. We report on an early prototype, SCOLA which allows the realization of powerful service chain-enhanced abstractions.

Tuesday, 13. October 2015

A Primer on IPv4 Scarcity
Speaker:
Philipp Richter
Type:
Project group meeting (Projektgruppentreffen, PGT)
Time:
13 October 2015
15:00 - 16:00
Place:
MAR 4.033
Abstract:

Content and services which are offered for free on the Internet are primarily monetized through online advertisement. This business model relies on the implicit agreement between content providers and users where viewing ads is the price for the "free'' content. This status quo is not acceptable to all users, however, as manifested by the rise of ad-blocking plugins which are available for all popular Web browsers. Indeed, ad-blockers have the potential to substantially disrupt the widely established business model of "free'' content---currently one of the core elements on which the Web is built.

In this work, we shed light on how users interact with ads. We show how to leverage the functionality of AdBlock Plus, one of the most popular ad-blockers to identify ad traffic from passive network measurements. We complement previous work, which focuses on active measurements, by characterizing ad-traffic in the wild, i.e., as seen in a residential broadband network of a major European ISP. Finally, we assess the prevalence of ad-blockers in this particular network and discuss possible implications for content providers and ISPs.

Tuesday, 06. October 2015

Improving Video Streaming QoE Through Multi Access Policies
Speaker:
Patrick Kutter (TU Berlin master student)
Type:
Masterstudent final talk
Time:
06 October 2015
17:00–18:00
Place:
MAR 4.033
Abstract:
Nowadays video streaming has become an important factor in the Internet.

With the standardization of Dynamic Adaptive Streaming over HTTP (DASH) the industry now has a common standard to develop and improve applications. The perceived subjective quality of a video stream however depends on several factors that have yet to be evaluated with client side applications for DASH. This thesis suggests the use of multiple network interfaces in combination with DASH to improve video streaming. However many existing solutions to manage multiple interfaces rely on static configurations for all connections. In this thesis the concept of Socket Intents is used. This concept takes the application specific needs into account by providing an extended Socket API that allows applications to transmit additional information in form of Intents when creating sockets. This concept has already been incorporated into the Multi Access Client (MUACC) framework. On top MUACC a new Socketconnect API has been suggested in cooperation with the maintainers of the framework. This new API has then been incorporated into MUACC by the maintainers. In thesis the corresponding policy for the new API has been developed, which allows the use of category and file size Intents. The new API then has been incorporated into the GPAC multimedia player in order to conduct a first evaluation of Quality of Experience (QoE) aspects by analyzing the application behavior in an emulated environment. This thesis finds that even with a very limited use of Intents the new approach allows a flexible distribution of network interfaces without negatively impacting QoE.

Tuesday, 22. September 2015

Joint Optimization of Concurrent and Consistent SDN Policy Updates
Speaker:
Felix Widmaier (TU Berlin master student)
Type:
Masterstudent final talk
Time:
22 September 2015
14:00–15:00
Place:
MAR 4.033
Abstract:
Today’s networks become increasingly dynamic. The demands for adequate load balancing, high utilization, and resilience against failures of hardware makes fine-grained managing and control of the network traffic necessary. Software Defined Networks (SDNs) can provide a high amount of flexibility due to their programmability and logical centralized view. Using this flexibility results in frequent policy updates. Executing an update will in this context be regarded as the implementation of a new rule on each corresponding switch from the source to the destination of one policy. As many routing problems can be solved efficiently in an SDN, due to the logical centralization of the controller, it is still necessary to take care of consistency properties in transient states during a policy update. The guarantee of loop-freedom is considered as a fundamental consistency property for networks and, therefore, forms the basis of all policy updates in this work.

Naturally, there are several coincident policy updates at a time, which have to be accomplished. To safe resources it is advantageous to combine these policy updates and, thus, to send less messages, i.e. to reduce the number of touches of the switches. In a theoretical analysis the Multi-Touch-Lock is presented, which prevents that a perfect solution, namely that every switch can be updated with only one touch, always exists. Furthermore, it is desirable to have an optimal solution which guarantees a small need of resources. Because of this, a program to find such a solution is implemented in OscaR, which is a toolkit to solve Operations Research problems. Unfortunately, this program cannot serve to schedule updates in a real life network, since policy updates have to be fast, which means that not only the implementation of an update has to have a short duration, but also the calculation of its schedule.

For this matter different heuristics are developed which focus on the reduction of the number of touches. They are compared to fast state of the art algorithms, which dynamically distribute updates between several rounds. In doing so, it is shown that the developed Single-Touch-Greedy heuristic can reduce the number of touches even more efficiently than fast up-to-date algorithms, which focus on the number of update rounds and are combined with a simple aligning algorithm. On the downside, Single-Touch-Greedy needs more rounds than the up-to-date algorithms.

Tuesday, 15. September 2015

An overview of Big Data analytics and the Apache Flink platform
Speaker:
Jon Crowcroft, University of Cambridge
Type:
Talk
Time:
15 September 2015
10:00–11:00
Place:
MAR 4.033
Abstract:
Much work in Data Center Networking has been about the need for speed. Research, and development, have concentrated on raw capacity, in transmission, switching, topology management. There are occasional bursts of work addressing a particular problem (TCP Incast, Outcast, load balancing) but in general the work proceeds along the same path, which is quantitatively important, but has not changed the nature of the data center network qualitatively for some time.

In this talk, I will discuss three pieces of work that we have been carrying out in Cambridge to address three different aspects of application needs directly.

Firstly, we have devised an extremely simple scheme, Qjjump, to provided hard bounded latency in the network, which is important for applications whose processing is typically "round trip time" bottlenecked (e.g. whose next step is determined by results of an RPC), but also can be used for more interesting services. The QJump system is simpler than any of the approaches for bounding latency that we have seen in the literature to date, and relies on the observation that we know a lot about the traffic sources and traffic matrix in a data center, and can reasonably rely on the absence of mis-use of the scheme, since the data center is, to some extent, a cooperative environment, and is, in any case, managed by a single organisation who can detect and remove mis-behaving applications.

Secondly, the Qjump system allows the possibility to provide very high probability, low latency, failure detectors. This means that distributed fault-tolerant applications (for example) that require this to complete and majority consensus algorithm, can now avoid the fate of the CAP theorem. The possibility to provide consistency, availability and partition tolerance is useful in fault tolerant computing in general, but its immediate application in data center networks could be to provide a simple way for SDN to update openflow rules simultaneously across the network, in the presence of faults.

Thirdly, we go further, and move some application code into switch processing resources within the Data Center Network. For applications that generate pathological traffic patterns that could disturb the admission control system (shuffle phase of map/reduce) this can help reduce the problem considerably, but many other simple stages of applications are also possible. Part of the challenge here is to cope with the switch processor heterogeneity.
Bio:
Jon Crowcroft is the Marconi Professor of Networked Systems in the Computer Laboratory, of the University of Cambridge. Prior to that he was professor of networked systems at UCL in the Computer Science Department. He has supervised over 45 PhD students and over 150 Masters students.

He is a Fellow of the ACM, a Fellow of the British Computer Society and a Fellow of the IEE and a Fellow of the Royal Academy of Engineering, as well as a Fellow of the IEEE. He was a member of the IAB 96-02, and went to the first 50 IETF meetings; was general chair for the ACM SIGCOMM 95-99; is recipient of Sigcomm Award in 2009. He is the Principle Investigator in the Computer Lab for the EU Social Networks project, the EPSRC funded Horizon Digital Economy project, hubbed at Nottingham, the EPSRC funded project on federated sensor nets project FRESNEL, in collaboration with Oxford; and a new 5-year project towards a Carbon Neutral Internet with Leeds.

Jon's research interests include Communications, Multimedia and Social Systems, especially Internet related.

Wednesday, 09. September 2015

Effectiveness of Network resilience in times of Hypergiants
Speaker:
Soumya Parida (TU Berlin master student)
Type:
Masterstudent introductory talk
Time:
09 September 2015
11:00–12:00
Place:
MAR 4.033
Abstract:
The Internet is built upon the principle of redundancy and failover. Especially routing and interconnection between different Autonomous Systems (AS) are highly redundant. But with the Internet becoming a mainstream product about a decade ago, the network redundancy alone has become secondary to content and service availability. With the emergence of Hypergiants, such as Google, Facebook, etc., more and more content is reliant on their presence.

In order to quantify the impact of the Hypergiants on content availability an overview on how prevelant these services are is needed. To this end, a web-crawler capable of not only downloading webpages, but also measure their delivery time, the presence of Hypergiant infrastructures in webpages, the usage of tracking cookies and the recording of all DNS messages is to be implemented. Furthermore, secondary webpages, i.e. links from the main page, are also to be crawled and the same information to be gathered.

Once the crawling is done, the data is going to be analyzed towards the prevalence of Hypergiants. Each website it to be classified by how many Hypergiants are present on each website, giving an overview on how far the reach of Hypergiants in todays Internet actually reaches. Finally, a first look on the impact of Hypergiant failure on content delivery it to be complied from the gathered data as an outlook on future work.

Tuesday, 01. September 2015

Incremental Parallel and Distributed Systems
Speaker:
Pramod Bhatotia

Type:
Talk
Time:
01 September 2015
10:00–11:00
Place:
MAR 4.033
Abstract:
Many applications exhibit the property that small, localized changes to the input result in only minor changes to the output.
Incremental computations seek to exploit this observation to speed-up successive runs of an application by tracking control and data dependencies, and re-executing only those parts of the computation that are affected by a given input change. To realize the benefits of incremental computation, researchers and practitioners are developing new systems where the programmer can provide efficient update mechanisms for changing application data. Unfortunately, most of the existing solutions are limiting because they either depart from existing programming models, or require programmers to devise an incremental update mechanism (or a dynamic algorithm) on a per-application basis.

The high-level goal of my research is to enable practical, automatic, and efficient incremental computations in parallel and distributed systems. My approach neither requires a radical departure from current models of programming nor complex, application-specific dynamic algorithms.

In this talk, I will first present a high-level description of my research. Thereafter, as a concrete example I will describe my recent project “iThreads”. iThreads is a POSIX-compliant threading library to support parallel incremental computation targeting unmodified C/C++ multithreaded programs. The iThreads library can be used as a drop-in replacement for the pthreads library, making it trivial to obtain the benefits of incremental computation by a simple exchange of libraries linked, without even recompiling the application code. To achieve this result, we design our algorithms and implementation to operate at the compiled binary code level by leveraging operating system-specific mechanisms. Our design choice of iThreads tries to strike a balance between efficiency of incremental computation and transparency for unmodified C/C++ pthreads-based programs. Our evaluation on a multicore platform using benchmarks from the PARSEC and Phoenix suites shows significant performance improvements for the majority of applications.

Thursday, 27. August 2015

Policy-compliant fast failover in Openflow networks
Speaker:
Tobias Martin (TU Berlin master student)
Type:
Masterstudent introductory talk
Time:
27 August 2015
14:00–15:00
Place:
MAR 4.033

Friday, 14. August 2015

Implementing and comparing new scoring techniques in the Cassandra client Astyanax
Speaker:
David Herzog (TU Berlin bachelor student)
Type:
Bachelorstudent final talk
Time:
11 August 2015
16:00–17:00
Place:
MAR 4.033
Abstract:
Obtaining low latencies for read requests in distributed storage systems has become important but difficult to achieve. Especially keeping the tail of latencies low is crucial for the performance of distributed systems. In this thesis we review the Cassandra client Astyanax developed by Netflix and illustrate its different replica selection mechanism.

In addition we discuss the recently developed replica selection mechanism C3 developed by Lalith Suresh et al. We implemented C3 inside Astyanax and performed experiments in our testbed to explore its performance under different conditions. Through our experiments we learned that our implementation of C3 in Astyanax in some settings of Cassandra has the lowest tail of latencies. We also discovered that the load on our Cassandra nodes can have significant impact on the performance of the replica selection mechanism of Astyanax. However, our results are only an indication of how Astyanax performs in a real world scenario since our experiments were performed with a small homogeneous Cassandra cluster in an non typical environment.

Tuesday, 11. August 2015

A Primer on IPv4 Scarcity
Speaker:
Philipp Richter
Type:
Project group meeting (Projektgruppentreffen, PGT)
Time:
11 August 2015
15:00 - 16:00
Place:
MAR 4.033
Abstract:

With the ongoing exhaustion of free address pools at the registries serving the global demand for IPv4 address space, scarcity has become reality. Networks in need of address space can no longer get more address allocations from their respective registries.

In this work we frame the fundamentals of the IPv4 address exhaustion phenomena and connected issues. We elab- orate on how the current ecosystem of IPv4 address space has evolved since the standardization of IPv4, leading to the rather complex and opaque scenario we face today. We outline the evolution in address space management as well as address space use patterns, identifying key factors of the scarcity issues. We characterize the possible solution space to overcome these issues and open the perspective of address blocks as virtual resources, which involves issues such as differentiation between address blocks, the need for resource certification, and issues arising when transferring address space between networks.

Friday, 31. July 2015

On the Suitability of Traceroute RTTs to Measure Geographic Distances
Speaker:
Stefan Gerganov (TU Berlin master student)
Type:
Masterstudent final talk
Time:
31 July 2015
11:00–12:00
Place:
MAR 4.033
Abstract:
Delay measurements are an integral part of researching Internet structure and behavior. They are often collected using an implementation of the classical Van Jacobson traceroute, generally applied for acquiring information about topology and latencies in the form of round-trip times. The advantage of traceroute over ping is that it provides data not only for the probed end-host, but also for all intermediate IP hops. Thus, it allows for latency characteristics to be studied along the entire path a packet traverses from the source to the destination. This can be particularly beneficial when looking at properties, such as the geographic location of the routing infrastructure itself in contrast to pure end-host latency measurements. In view of this fact, geolocation, the mapping of IP addresses to geographic locations, is the example use case chosen for this work.

There are, however, numerous pitfalls associated with applying traceroute measurements to real-world use cases. This thesis aims to uncover some of the less-often explored and thus usually overlooked pitfalls when attempting to infer the location of Internet routing infrastructure from the output of traceroute. The approach studies the suitability and increases usability of the various metrics that traceroute measurements yield. It primarily builds upon speed-of-light latency constraints using traceroute data from the CAIDA Ark infrastructure, a globally distributed set of probing devices which continuously traceroute the entire IPv4 address space.

The study begins with a careful comparison between traceroute and ping in terms of delay differences they produce for the same locations, assessing the relationship between speed of light and latencies from both tools. After this step, the suitability of traceroute measurements to infer geographic locations is evaluated by using pure round-trip times, as well as differences between latency pairs, also referred to as delta round-trip times. Hereby, several non-obvious issues with traceroute latency values are identified, stemming from ICMP reply type-code combinations, as well as MPLS tunnels. The former intro- duce unexpectedly large delays, which skew median latencies towards the thousands of milliseconds, while the latter often produces negative values when using differences be- tween round-trip times.

Tuesday, 07. July 2015

Implementation and Evaluation of Multi-Access Policies for MPTCP Path Management in User-Space
Speaker:
Mirko Palmer (TU Berlin master student)
Type:
Masterstudent final talk
Time:
07 July 2015
15:00–16:00
Place:
MAR 4.033
Abstract:
Today, most mobile end-user devices have multiple network interfaces, but usually only use one at a time. There are several approaches to improving the handling of multiple interfaces and to enable their simultaneous usage. One of these is Multi-Path TCP (MPTCP), which makes it possible to split up TCP streams into multiple sub-flows and therefore enables applications to transparently use all available interfaces at once. Another approach is the multi-access management framework (MAM), developed here at FG-INET. It incorporates Socket Intents to give applications the ability to make an informed decision regarding which single interface is best suited for a specific scenario.
Because the current Linux reference implementation of MPTCP is not capable of making informed decisions about which interface to use, this thesis aimed to overcome this limitation and to combine the approaches of MPTCP and MAM. Instead of always using all available interfaces, MPTCP was modified to be controlled by MPTCP-aware policies within the MAM running in user-space. Thus, we gained the benefit of utilizing multiple interfaces combined with the capability to intelligently choose the best combination by letting applications take part in the decision.
By measuring the time simulated HTTP downloads of different web pages required, the performance of the combined approach was evaluated in several scenarios and found to be superior in some cases.

Tuesday, 07. July 2015

Investigating a reordering robust TCP for datacenter and storage cluster use
Speaker:
Puneeth Nanjundaswamy (TU Berlin master student)
Type:
Masterstudent final talk
Time:
07 July 2015
14:00–15:00
Place:
MAR 4.033
Abstract:
Transmission control protocol (TCP) is a highly reliable host-to-host protocol which ensures reliable transfer of data over unreliable internet. TCP must recover from damaged, lost, duplicated and out-of-order data. TCP sender retransmits a data segment upon receiving three duplicate acknowledgements(ACK) from the receiver. Since TCP does not know whether a duplicate ACK is caused by a lost segment or just a reordering of segments, it waits for 3 duplicate ACKs to be received. It is assumed that if there is just a reordering of the segments, there will be only one or two duplicate ACKs before the reordered segment is processed, which will then generate a new ACK.

However, the above assumption is not valid any more as numerous studies in the wild internet have found that packet reordering is not rare and that waiting for three duplicate ACKs might not be sufficient to disambiguate packet reordering and packet loss. Retransmission of data segments also results in the decrease of the sender's throughput. Thus an efficient reordering detection and reaction mechanism is needed to avoid spurious retransmissions. Avoiding spurious retransmissions reduces the unnecessary decrease in the sender's throughput, there by improving the performance of TCP.

This thesis describes packet reordering and its metrics in detail.
Furthermore, it studies the performance impacts of packet reordering on TCP. It then investigates the various approaches by which TCP can be made robust against packet reordering with focus on the Linux's implementation, Non congestion to robustness (NCR) and adaptive NCR algorithm (aNCR) of detecting and handling packet reordering. The algorithms are implemented in the Linux kernel and evaluated under various conditions of packet reordering and under multiple traffic conditions to determine the best performing algorithm under all scenarios.

Tuesday, 30. June 2015

Fun With Integer Linear Programs
Speaker:
Matthias Rost
Type:
Project group meeting (Projektgruppentreffen, PGT)
Time:
30 June 2015
15:00 - 16:00
Place:
MAR 4.033
Abstract:
(Integer) Linear Programming is a tool that has  been around for more than half a century now. It has been proven useful in nearly every aspect of operations research including but not limited to:
  • packing dish washers
  • computing the cheapest diet for soldiers / PhD students
  • scheduling in chemical processing plants
  • computing routings of any sort (VLSI, optical, IP-based...)
In this talk I will present some applications, outline solution techniques for (Integer) Linear Programs and report on some of our insights.

Disclaimer: The talk will be extremely non-technical. Everybody is welcome.

Tuesday, 16. June 2015

Cloud Slam: Thirty Minutes, Three Talks, You Judge
Speaker:
James Kempf
Type:
Project group meeting (Projektgruppentreffen, PGT)
Time:
16 June 2015
15:00 - 16:00
Place:
MAR 4.033
Abstract:
In 2014, I finished up a 4 year program on networking as a service (Cloud + SDN). Moving forward, I've been thinking about three new topics:

  1. Spot pricing for operator bandwidth
  2. Which technique is better for cloud resource management: bid or offer?
  3. Bayesian causality and cloud policy
In this presentation, I'll give three short micro-talks about these topics in the popular Science Slam format. This is not finished work but preliminary thinking about what could come next. At the end of the talks, participants will get to judge which topic or topics they think is/are the most promising.

Tuesday, 09. June 2015

Scheduling Loop-free Network Updates: It's Good to Relax!
Speaker:
Arne Ludwig
Type:
Project group meeting (Projektgruppentreffen, PGT)
Time:
09 June 2015
15:00 - 16:00
Place:
MAR 4.033
Abstract:
We consider the problem of updating arbitrary routes in a software-defined network in a (transiently) loop-free manner. We are interested in fast network updates, i.e., in schedules which minimize the number of interactions (i.e., rounds) between the controller and the network nodes. We first prove that this problem is difficult in general: The problem of deciding whether a k-round schedule exists is NP-complete already for k = 3, and there are problem instances requiring Ω(n) rounds, where n is the network size. Given these negative results, we introduce an attractive, relaxed notion of loop-freedom. We prove that O(log n)-round relaxed loop-free schedules always exist, and can also be computed efficiently..

Tuesday, 26. May 2015

Managing peering link congestion in times of volatile CDN traffic
Speaker:
Mama Ely (TU Berlin master student)
Type:
Masterstudent final talk
Time:
26 May 2015
15:00–16:00
Place:
MAR 4.033
Abstract:
Internet content delivery via Content Delivery Networks (CDNs) has transformed traffic patterns for Internet Service Providers (ISPs). Thus, long term predictions are becoming harder as the traffic is getting more volatile due to large distributed deployment and dynamic server selection. While ISPs usually manage this in their own Autonomous Systems (ASes), the links between different ASes pose an increasing challenge.

This is in no small part due to content delivery from CDNs usually having heavy inbound traffic, i.e. the major traffic flows from the CDN server to the ISP. Engineering this traffic from the ISP side is at fundamental odds with the design of Internet routing, as this is only forward based. Thus, by the time an ISP wishing to engineer traffic on its inter-AS links has control of the traffic, the crucial inter-AS link has already been traversed. Classic solutions to tackle this problem are based on tuning the Border Gateway Protocol (BGP) used for exchanging Information on inter-AS links, i.e. route-aggregation, and tuning BGP metrics, such as communities. However, all information conveyed this way are non-binding and only reach the neighboring AS.

In this thesis, we propose to open a direct communication channel from the ISP to the CDN, i.e. the sending party, via the Provider-aided Distance Information System (PaDIS). Here, the ISPs gain the ability to communicate their preferences on where to send data from, negating the reverse path problem and leading to the ability to choose inter-AS links as a forwarding decision.

Tuesday, 26. May 2015

Investigating Rate Control in the Context of Replica Selection to reduce Tail Latencies
Speaker:
Tobias Brandt (TU Berlin master student)
Type:
Master student introductory talk
Time:
26 May 2015
14:00–15:00
Place:
MAR 4.033
Abstract:
The recently proposed C3 system uses adaptive replica selection for reducing tail-latencies in the context of replicated and partitioned data-stores.
In order to ensure that a set of clients do not saturate individual servers in the system, C3 uses distributed rate-control, wherein clients adapt their sending rates to servers in a decentralized fashion. The original C3 system used an adaptation of the TCP-CUBIC congestion control algorithm in order to perform distributed rate-control. The goal of this thesis is to investigate alternative rate adaptation mechanisms for the C3 system.
In particular, we will focus on the i) latency based rate-adaptation mechanisms of the PARDA and PISCES systems (inspired by FAST-TCP), ii) the BIC mechanism used by the PULSAR system, as well as iii) a simple AIMD approach.

Tuesday, 19. May 2015

Masterstudent final talk
Speaker:
Bogdan Ciobotaru (TU Berlin master student)
Type:
Masterstudent final talk
Time:
29 May 2015
15:00–16:00
Place:
MAR 4.033
Abstract:
MiddleBoxes (MBs) or specialized appliances are a crucial part of today's enterprise networks in order to improve the performance and security, e.g., through firewalls, NAT or intrusion detection systems. However, today's middleboxes are expensive, their management is cumbersome and are responsible for adding new failure modes to a network. Moreover, MB often come as hardware appliances making the integration into existing network architectures difficult. Hence, the ETSI Network Functions Virtualization (NFV) initiative aims to reduce deployment, space and management costs by implementing network functions in software running on commodity hardware.

By introducing programmability into the network, the software-defined networking (SDN) paradigm greatly improves the flexibility of network management and operation, and enables innovative new network services. However, the interaction between NFV and SDN poses many challenges, e.g., the client's connection state of a stateful firewall MB is not moved to a new MB in the presence of client mobility. Our system, called vMB, introduces the notion of per-client virtual middle boxes (vMBs), to a abstract and decouple user-specific state. Specifically, vMBs are used to fully reap the virtualization benefits: the handling of vMBs is important to guarantee the decoupling of the per-client middlebox state and the inner workings of the middlebox from the physical instance. In this talk, we present the system and two vMB driver implementations for the Linux stageful firewall, based on: 1) the netfilter conntrack tool, and 2) the Linux Kernel netlink interface.

Wednesday, 22. April 2015

Mobile Core Network Virtualization: An Optimization Model for Virtual Network Function Placement and Service Chaining
Speaker:
Andreas Baumgartner (Prof Bauschert’s student)
Type:
Talk
Time:
22 April 2015
10:00–11:00
Place:
MAR 4.033
Abstract:
Der Vortrag wird mit Bezug zum VINO Projekt gehalten. Die anschließende Diskussion wird über den im Vortrag vorgestellten Optimierungsansatz und alternative Ansätze, sowie Möglichkeiten, die Rechenzeit zu reduzieren > Online Optimierung, gehen.

Tuesday, 21. April 2015

Graph Signals and Filters
Speaker:
Andreas Loukas
Type:
Project group meeting (Projektgruppentreffen, PGT)
Time:
21 April 2015
15:00
Place:
MAR 4.033
Abstract:
We have recently seen a surge of research focusing on the processing of data occurring on graphs. The emerging field of signal processing on graphs focuses on the extension of classical signal processing techniques to the graph setting. Arguably, the greatest breakthrough of the field has been the extension of the Fourier Transform from time signals and images to graph signals, i.e., signals defined on the nodes of irregular graphs. By providing a graph-specific definition of frequency, the graph Fourier transform enables us to the design graph filters, analogues of classical filters but suitable for the graph setting. In this talk, I will give a short introduction to the field and I will review recent advances on the design of distributed graph filters.

Tuesday, 10. February 2015

C3: Cutting Tail Latency in Cloud Data Stores via Adaptive Replica Selection
Speaker:
Lalith Suresh Puthalath
Type:
Project group meeting (Projektgruppentreffen, PGT)
Time:
31 March 2015
14:00–15:00
Place:
MAR 4.033
Abstract:
Achieving predictable performance is critical for many distributed applications, yet difficult to achieve due to many factors that skew the tail of the latency distribution even in well-provisioned systems.

In this talk, we will present the fundamental challenges involved in designing a replica selection scheme that is robust in the face of performance fluctuations across servers. We illustrate these challenges through performance evaluations of the Cassandra distributed database on Amazon EC2. We then present the design and implementation of an adaptive replica selection mechanism, C3, that is robust to performance variability in the environment. We demonstrate C3's effectiveness in reducing the latency tail and improving throughput through extensive evaluations on Amazon EC2 and through simulations. Our results show that C3 significantly improves the latencies along the mean, median, and tail (up to 3 times improvement at the 99.9th percentile) and provides higher system throughput.

Monday, 16. March 2015

IoT In-Network Aggregation
Speaker:
Damien Foucard, Niklas Semmler
Type:
Project group meeting (Projektgruppentreffen, PGT)
Time:
16 March 2015
14:00–15:00
Place:
MAR 4.033
Abstract:
The topic of our talk will be the estimation of the top items appearing in multiple data streams with limited communication overhead..

Tuesday, 17. February 2015

On the Potentials of Extending Conventional Measurement Tools for Network Deployments using ALTO
Speaker:
Johannes Schütze (TU Berlin master student)
Type:
Masterstudent final talk
Time:
17 February 2015
14:00–15:00
Place:
MAR 4.033
Abstract:
Nowadays it is a common practice for Internet Service Providers (ISPs) to hide their network topology and configuration. The reason behind it is that this knowledge can be used to refine malicious attacks on the network or reveal the operators business strategy. However the knowledge about the network of an ISP can also be used to benefit its users and partners. For example, Content Distribution Networks (CDNs) and Peer-to-Peer (P2P) clients rely on knowledge of the network topology to appropriately select their peers and map user to close by servers. However, today their only choice is to do active Internet measurement to improve these services. This is a huge overhead for the P2P clients and CDNs and its results are volatile. The Application Layer Traffic Optimization protocol (ALTO) is a new technology that aims to supply topological information to improve the end-users experience, while keeping the topology hidden. In short, ALTO it designed as a broker between the ISP and the users of the network it describes, aiming at aligning the interests of both parties for improved performance.

This thesis evaluates ALTO in conjunction with Internet measurement and gives insight into the efficiency of hiding topology information, potentials and configuration of ALTO. Multiple ALTO aggregation levels are defined and evaluated network based on real ISP data as well as artificial network structures to show how ALTO is used to provide better end-user experience, while successfully minimizing the amount of revealed topology information. For this purpose a framework is implemented that simulates ALTO and Traceroute to generate this statistical information.

Tuesday, 10. February 2015

myFi: Enabling Fine Grained Control in Home Networks
Speaker:
Julius Schulz-Zander
Type:
Project group meeting (Projektgruppentreffen, PGT)
Time:
10 February 2015
15:00–16:00
Place:
MAR 4.033
Abstract:
The quickly growing demand for WiFi networks and the numerous application-specific requirements stand in stark contrast to today’s inflexible management and operation of home networks. In this paper, we present myFi, a novel WiFi architecture based on off-the-shelf access points. myFi allows to make a more efficient use of the wireless spectrum, by introducing service differentiation and fine-grained transmission control. myFi uses per-client virtual access points and per-client virtual middleboxes, to render network functions more flexible and support seamless migration as well as user mobility. myFi can also be used to out-source control over network slices (e.g. Hotspots) to an ISP. We report on a prototype implementation of myFi and show that it allows to effectively prioritize different applications.

Thursday, 05. February 2015

Measuring the effects of IPv4 address exhaustion on allocation and routing dynamics
Speaker:
Goncalo Morais (master student)
Type:
Masterstudent final talk
Time:
05 February 2015
11:00–12:00
Place:
MAR 4.033
Abstract:
In 2011 the Internet Assigned Number Authority (IANA), allocated the last block of IPv4 addresses to the Regional Internet Registries (RIRs). The RIRs in turn reached a critical level of available IPv4 addresses for further allocation, which indicates a state of near depletion of IPv4 addresses. The fact that we are facing IPv4 address exhaustion makes IPv4 addresses an increasingly scarce resource. Hence, we observe changes on IPv4 allocation policies by the Regional Internet Registries (RIRs) and the appearance of IPv4 transfer markets. Using BGP data from RouteViews and the allocation files provided by the RIRs, we propose to investigate the effects that this is having on the allocation and routing dynamics. Hereby we pay particular attention to how origins of prefixes have been changing, how long it takes for a prefix to be seen in routing after it has been allocated and the routing status of prefixes. We will also infer about the number of ASes that advertise allocated blocks of addresses and how this has changed along the years. We will finish by having a look at how some prefixes are being documented in registry mechanisms such as the Regional Internet Registries and the Internet Routing Registries. Our study shows a change on the allocation and routing dynamics of newly allocated prefixes, but also the so called "Legacy" allocations show a change in their routing dynamics and as a result the governance of address space should adapt to this new situation.

Tuesday, 03. February 2015

Optimizing SDN Updates With Transient Consistency Guarantees
Speaker:
Damien Foucard (master student)
Type:
Masterstudent final talk
Time:
03 February 2015
14:00–15:00
Place:
MAR 4.033
Abstract:
Software-defined networking (SDN) is a paradigm rich of opportunities. The administrator defines a simple global policy in the control plane and the controller automatically translates it into routing tables of each switch, thus shielding the administrator from the complexity of vendor-specific switch configurations. Hence, the administrator simply needs to request the controller to enforce a particular constraint, and it will automatically compute a new policy in keeping with this constraint, and updates the network to enforce this policy. Unfortunately, while each singular policy is consistent with specified constraints, the network can remain for an arbitrary large amount of time in a transient state violating them if no specific measure is taken. To avoid this problem, the controller should make sure that the network can only go through transient states conforming to specified constraints, which requires scheduling some updates before others. However, this possibly implies that many switches must wait for others to implement the new policy. A shortest possible update schedule – reducing the latency before the new policy is enforced to the minimum while respecting specified constraints– appears then desirable but computing it induces a high computational cost, potentially not affordable in real time. We consider two constraints: loop-freedom (no packet should loop) and waypoint enforcement (all packets must traverse a middlebox). Loop-freedom is desirable because, although packets caught in loops are usually automatically removed (e.g., with TTL), they represent an undesirable load. Waypoint enforcement is a key feature for Network Function Virtualization and is of particular relevance in security-critical environment, where allowing packets to bypass a middlebox even for a short time already represents an unacceptable risk. An ideal algorithm for computing an update schedule would be easy to compute (minimized computational cost), update the network quickly (routers implementing new policy in minimal time), and offer strong consistency guarantees (maximized number of constraints ensured). Unfortunately, as we show in this work, it appears that this Graal does not exist, and favoring one dimension comes at the expense of another. Hence, balances need to be struck. Accordingly, we consider a variety of settings, favoring in turn one dimension or the other, and in each case present optimized algorithms and highlight caveats. In addition, we implement a framework in which we simulate our algorithms and measure their performance, and present our results. The framework’s modular design allowing for an easy expansion.

Monday, 02. February 2015

Pretty Bad Privacy: Pitfalls of DNS Encryption
Speaker:
Haya Schulman
Type:
Talk
Time:
02 February 2015
14:00–15:00
Place:
MAR 4.033
Abstract:
As awareness for privacy of Domain Name System (DNS) is increasing, a number of mechanisms for encryption of DNS packets were proposed. We study the prominent defences, focusing on the privacy guarantees, interoperability with the DNS infrastructure, and the efficiency overhead. In particular:
  • We explore dependencies in DNS and show techniques that utilise DNS specific side channel leaks allowing to infer information about the target domain in an encrypted DNS packet.
  • We examine common DNS servers configurations and show that the proposals are expected to encounter deployment obstacles with (at least) $38\%$ of 50K-top Alexa domains and (at least) $12\%$ of the top-level domains (TLDs), and will disrupt the DNS functionality and availability for clients. We also show the implication of these configurations on adoption of DNSSEC.
  • We show that due to the non-interoperability with the caches, the proposals for end-to-end encryption may have a prohibitive traffic overhead on the name servers.
Our work indicates that further study may be required to adjust the proposals to stand up to their security guarantees, and to make them suitable for the common servers' configurations in the DNS infrastructure.

Tuesday, 27. January 2015

An overview of Big Data analytics and the Apache Flink platform
Speaker:
Asterios Katsifodimos
Type:
Project group meeting (Projektgruppentreffen, PGT)
Time:
27 January 2015
15:00–16:00
Place:
MAR 4.033
Abstract:
Apache Flink is a software stack for complex big data analytics, which originates from the Stratosphere research project coordinated at TU Berlin. Flink covers a variety of use cases, such as data warehousing, information extraction and integration, data cleansing, graph analysis, and statistical analysis applications. Flink brings together a unique set of features from database systems and MapReduce-like technology that enables expressive, easy, and efficient programming for the development of large-scale analytical applications. Flink features include in situ data processing, multiple language APIs (e.g., for Java and Scala), treatment of user-defined functions (UDFs) as first-class citizens, automatic program parallelization and optimization, support for iterative programs, and a scalable and efficient execution engine. In this talk, I will present an overview of the Stratosphere vision, the Flink platform as well as current and future research activities.
Bio:
Asterios Katsifodimos is a Postdoctoral Researcher co-leading the Stratosphere research project at the Technische Universität Berlin. He received his PhD in Computer Science in 2013 from INRIA Saclay and Université Paris-Sud under the supervision of Ioana Manolescu. His thesis focused on materialized view-based techniques for the management of Web Data. Asterios was a member of the High Performance Computing Lab at the University of Cyprus, where he obtained his B.Sc. and M.Sc. degrees in Computer Science. His research interests include query processing, optimization, and massively parallel data analysis.

Zusatzinformationen / Extras

Quick Access:

Schnellnavigation zur Seite über Nummerneingabe

Auxiliary Functions