CoNEXT'16 Session 11: Measurements and diagnosis 2

1. "Measuring Latency Variation in the Internet", (short paper) 
Toke Høiland-Jørgensen (Karlstad University), Bengt Ahlgren (SICS), Per Hurtig (Karlstad University), and Anna Brunstrom (Karlstad University)

Internet latency has large variations that can come from many sources. Although the network throughput has been largely increased over the past few years, both the minimum Internet latency and the latency variation have not been improved over time. This works focuses on measurement the latency variation experienced by a client by utilizing the extensive publicly available dataset from the Measurement Lab Network Diagnostic Tool (NDT) with a packet capture from within a service provider access network.

It is found out that there is significant regional differences in the datasets. For instance, Africa, Asia, Europe and America all have different RTT values. There are multiple sources that can cause such latency variations, including queueing delay along the path, delay acknowledgements, transmission delay, medium access delay, etc. One main source for Internet latency variation is the queueing delay along the path. The measurement results show that the queueing delay is a non-trivial number of instances and when the queueing delay occurs, it has significant impact on the network performance.

Q: In your experiments, where do you measure the latency?
A: We focused on the path from client to the server, in the TCP handshake delay measurements.

Q: The connection setup time, for example, the setup time for TCP connections, is critical. Have you look into the connection time to different servers such as Google, Facebook, etc.?
A: That should be useful measurements but we haven't done that in this work.

Q: An thoughts on extending this work to the end-to-end latency?
A: We definitely think about the metrics, but however, the current datasets and platform limit such measurements. Will look into it in the future.


2. "LossRadar: Fast Detection of Lost Packets in Data Center Networks",Yuliang Li (Yale University), Rui Miao (University of Southern California), Changhoon Kim (Barefoot Networks), and Minlan Yu (Yale University)

Packet loss diagnostics is very important in data centers. Packet losses come very common at large-scale data centers. Nowadays, it usually take a very long time for data centers to detect and locate packet losses. Packet losses can have high impact on data centers' performance: it will increase the latency, drop the throughput, and break the connection between clients and the servers. There is no effective ways today to track the root causes and there are many challenges in doing so. Unfortunately, existing monitoring tools that are generic in capturing all types of network events often fall short in capturing losses fast with enough details and low overhead.

Therefore, we design LossRadar, which is a system that can capture individual lost packets and their detailed information in the entire network on a fine time scale. A high level idea of this approach it that each switch mixes all the packets they see in a small data structure. The same packets in the data structure will cancel out and only the loss packets will remain. The traffic digests are built using Invertible Bloom Filter structure, which has a table associated with multiple hash functions. Each packet will be hashed by each hash function.

LossRadar is demonstrated to be easily implemented using P4 language. LossRadar is evaluated in simulations using 80 switches from 28 hosts with 10G links. Simulation results show that the memory usage is very low, for 10G traffic, LossRadar only needs less than 8kb to catch the losses. LossRadar is also compare to state-of-the-art approaches and LossRadar is shown to have lower memory, while achieving a good performance. The P4 code of LossRadar is also released on github.

Q: Can you explain a bit more about how the loss differentiation algorithm works?
A: The different types of losses are described in three dimensions and the bursty is quantified by the timestamp of each loss.

Q: Can LossRadar be extended to handle the packet loss at the end host (not in the switch)?
A: Yes it can! Just need to install the meter in the server to detect the loss, it's the same as detecting losses between the switches.

Q: Have you evaluated the performance overhead on the forwarding path?
A: It doesn't affect the data path of the packets.


3. "Demystifying and Puncturing the Inflated Delay in Smartphone-based WiFi Network Measurement", (short paper)
Weichao Li (The Hong Kong Polytechnic University), Daoyuan Wu (Singapore Management University), Rocky K. C. Chang (The Hong Kong Polytechnic University), and Ricky K. P. Mok (CAIDA/UCSD)

We have a lot of apps measuring the latency of the network on our mobile phones, but how about their accuracy? Are they doing exactly what they are supposed to do? This work present the measurements of network-level latency, instead of the popular user-level latency, and looks into the source from the kernel space instead of from the users. The two main contributions of this work is (i) energy-saving mechanisms - theSecure Digital Input Output (SDIO) bus sleeping and the IEEE 802.11 Power Saving Mode (PSM) - are the main sources of the delay inflation, and (ii) the design, implementation, and evaluation of AcuteMon, which is an Android app prototype run on unrooted phones and requiring no system customization, such as kernel recompilation and customized ROM.

Controlled experiments are produced on several phone models to collect the delay from the network and from the drivers. Experimental results achieve very accurate network RTT and show that the delay overheads can be largely mitigated by the proposed approach.

Q: All the experiments are under controlled environments, have you conducted experiments in real-world where there may be interfering WiFi transmissions?
A: We are actually doing that right now and we have also done experiments in the wild.

Q: Have you measured the power consumption of the measurement?
A: We haven't done so since we focused on measuring the delay. But we will look into it in the future.

Q: How about the delay in 3G/4G networks?
A: We have had results for 3G/4G networks in terms of delay measurements, but they are not in this paper.

CoNext 2016 Session 2 -- Wireless 1

1. EMPoWER Hybrid Networks: Exploiting Multiple Paths over Wireless and ElectRical Mediums 

Authors: Sébastien Henri (EPFL), Christina Vlachou (EPFL), Julien Herzen (Swisscom), and Patrick Thiran (EPFL)

Besides the  advances in mobile computing the high throughput demands cannot always be satisfied because many technologies co-exist (WiFi, power-line communications -PLC-, cellular) but neither  cooperate nor operate at their full capacity. The authors propose the architecture EMPoWER, a system that exploits simultaneously several potentially interfering mediums. EMPoWER operates at layer 2.5 between the MAC and IP layers, and combines routing (to find multiple concurrent routes) and congestion control (to efficiently balance traffic across the routes), by exploiting the rich diversity offered by the multiple networks.

Network aggregation solutions can improve the throughput but routing must be carefully assigned 

in order to avoid congestion in the intermediate nodes. Thus, the paper studies a global throughput optimization problem with several congesting flows for multi-path scenarios. EMPoWER can be used with any protocol, but this work focuses only on a Hybrid Network of WiFI and PLC. Power line communications (aka PLC) is communication through electrical wires, and actually WiFi and PLC can be aggregated in WiFi relays). 

EMPoWER aims to find the optimal rates, avoid congestion, combine wired and wireless paths and provide a multi path controller. Constraints of the optimization problem are: 1. Airtime demand, 2. Airtime must not exceed 100%, 3. Interference of WiFi and PLC links. The controller for the multi-path routing protocol focuses on the total throughput and optimizes a utility function globally. 

The authors simulated their proposed schema and they compared it against the back pressure schema which has been proven optimal. Thus, two schedulers were used: 1. Optimal Scheduler for getting the optimal Throughput, 2. Scheduler with constraints under EMPoWER.

For the numerical results, 1000 random topologies were simulated. EMPoWER was close to the optimal throughput. EMPoWER includes also an actual testbed implementation at layer 2.5, IEEE 905 compatible. Authors demonstrate experimental results where aggregation improve users' throughput. Finally, a comparison between the hybrid PLC and WiFi schema vs multi-channel WiFi is considered. EMPoWER outperforms in 75% of scenarios and can offer up to 10x improvement.

Q1: Does the experimentation included a calibration  phase for calculating the network capacity of the wireless (WiFi) and the wired (electrical lines) medium? Hoes does the capacity change over time due to environment changes (e.g. electrical devices interference etc)?

A1: There are training symbols/information in the header about the modulation schema of WiFi/PLC in order to calculate the capacity. Of course, interference indeed does change the capacity, but the EMPoWER controller can handle time varying capacity changes.

Q2: Could you provide details for the simulator technology?

A2: It was implemented in Matlab and it’s available online.

2. FlexRAN: A Flexible and Programmable Platform for Software Defined Radio Access networks

Authors: Xenofon Foukas (The University of Edinburgh), Navid Nikaei (Eurecom), Mohamed M. Kasse (The University of Edinburgh), Mahesh K. Marina (The University of Edinburgh), and Kimon Kontovasilis (NCSR Demokritos).

Current 4G technologies haven’t been designed to support Internet of Things (IoT), device-to-device (D2D) and machine-to-machine (M2M), therefore, 5G must include programmable network and radio operations, easily and quickly adaptable to the traffic requirements  (i.e.“softwarization” of each stage). Apart from software radio networking (SDN), next generation cellular networks will include software defined radio access network (SD-RAN), and current literature hasn’t yet demonstrated a concrete platform which handle the idiosyncrasies of RAN control.

This work develops an open source SD-RAN platform for experimentation, which support real time applications,  provides a modular SD-RAN design, and allows programmable network functions and programmable radio layer operations. The master controller is a top level orchestrator of the FlexRAN and provides an API for fetching mobile and network statistics from the various entities of the network (UEs, eNBs, management, statistic entities etc). Moreover, the controller allows priorities to network flows at each different component.

Logical separation of control (network) plane/operations and data plane are supported by FlexRAN. For example, downlink traffic scheduling and uplink traffic scheduling flows can be completed separated  Finally, the Virtual Subsystem Functions are being executed by a re-programable scheduler (i.e. different scheduling per type of traffic). Its hierarchical master-agent design controller architecture  is well suited  for real time RAN control operations while allows reprogrammability and reconfigurability and control delegation following Network Function Virtualization.

FlexRAN master controller has been implemented in C++ and the FlexRAN agent is written in C. The whole system has implemented over the OpenAirInterface (OAI) LTE platform. The experimentation with the system showed no difference in service quality compared to preexisting controller Vanilla  of OAI.

Experimental results show that even in the worst case configuration (i.e. running real time schedulers on top of master controller which constantly sends statistics per ms) FlexRAN could respond to the resulted load without any degradation of QoS. For 50 UEs (worst case scenario) stats reporting inside FlexRAN could be up to 100 mbps (intra-traffic communications between the controller and the Network Functions  is a crucial performance factor).

Q: How would be FlexRAN be able to handle adaptive video streaming?

A: FlexRAN is able to create a new virtual operation per a new policy (i.e.  a new requested bitrate), The Fair scheduling policy per node dynamically adjusts the bitrate so the user will experience a smoothly transition in his bitrate.

3. Mudra: User-friendly Fine-grained Gesture Recognition Using WiFi Signals.

Authors: Ouyang Zhang (Ohio State University) and Kannan Srinivasan (Ohio State University)

This work presents Mudra, a framework for fine grained recognition of hands’ gestures. The authors leverage pervasive WiFI signals  in order to detect extremely fine, subtle finger gestures which can be used to human-to-machine interaction or other type of control. This framework can enable over the air interaction (i.e. resolve screen limitation) and allow communication in contact-forbidden scenarios  (e.g. infectious medical task).Prior work’s limitations include the need of large scale space for the deployment of large antenna array methods and that is limited to coarser hands’ gestures.

Mudra uses a two-antenna receiver to detect and recognize finger gesture. It uses the signals received from one antenna to cancel the signal from the other. This

“cancellation” is extremely sensitive to and enables the detection of small variation in channel due to finger movements. Since Mudra decodes gestures with existing WiFi transmissions, Mudra enables gesture recognition without sacrificing WiFi transmission opportunities. Besides, Mudra is user-friendly with no need of user training. Mudra’s prototype implementation was done on a NI-based SDR platform and used COTS WiFi adapter. Experimental evaluation showed that the system can achieve 96% accuracy.

Q: Did the experimentation include other movements in the environment apart from the fingers’ gesture?

A: No, future work will investigate how other body’s movement could affect the system.

4. Passive Communication with Ambient Light

Authors:  Qing Wang (Delft University of Technology), Marco Zuniga (Delft University of Technology), and Domenico Giustiniano (IMDEA Networks Institute)

This work studies, implements and evaluates a Visible Light Communications system. The transmitters are LEDS (i.e. bits 0/1 are represented by a specific light wavelength), but the information is modulated by the environment, i.e. via reflection. The receiver of this communication schema is a photodiode or a camera (i.e. a device which is capable detecting light changes).

The key concept and the motivation around this short range communication schema using visible light spectrum, is the sustainability due to the pervasive nature of the light. Communication via visible ambient light offers several advantages such as high data rate, localization via the light illumination etc. However, ambient light cannot be easily controlled and there are many challenges in the design of this system.

The system consists of: (a) An emitter (any light source) (b) Surface with reflective material and (c) a receiver which is a light sensitive optical device such as a photodiode. The surface modulates and reflects the incoming light depending on the data source to be transmitted.

The channel capacity (i.e. the throughput) depends on several parameters such as the inter-symbol interference, the symbol wavelength etc. Thus, the throughput is time variant and moreover decreases exponential with the distance.

There are several system design challenges such as defining the reflection properties  and the dimensions of the modulating surface, the channel distortion from the moving vehicle  (i.e. time varying speed), packet collisions due to multi-path etc. The modulation/coding scheme uses high reflection coefficient material for bit 1 and vice versa for bit 0. Most importantly, after receiving the symbol the receiver can perform the decoding based on the received signal strength of reflected light. The system evaluation included outdoor evaluation with mobile object at 18 km/h .

Q1: How is your system compared to a barcode? 

A1: For decoding a barcode, a “camera” is being utilized as a receiver, the light source is constant (i.e. the transmitter is the barcode picture). The concept is different and it is a static scenario. The communication system in this paper considers a moving vehicle as the transmitter and the receiver.

Q2: How do you envision the generalization of the system? 

A2: Future directions include: More controllable as possible via “re-programmable” surface properties, embedding mobility, considering cloudy vs sunlight conditions etc.

CoNEXT 2016: Session 8 New Directions

Source Accountability with Domain-brokered Privacy 

Taeho Lee (Speaker), Christos Pappas, David Barrera, Pawel Szalachowski, Adrian Perrig 

There are two network design objectives: accountability and privacy. And they are conflicting    in network design.

This works tries to arrive a solution that both features are preserved.

The limitations of current best practice, APIP SIGCOMM’14, is two-fold: selective packet authentication and error reporting.

The speaker defines two objectives in a formal way: source accountability and communication privacy.

The solution has three components:

1/ ISP as accountability agent and privacy broker.

2/ Ephemeral ID: temporary network address given by the ISP

3/ Certificate for Ephemeral ID given from ISP to host 

Q: Privacy guarantee under analysis with temporal correlation?

A: (wasn’t able to catch)

Q: About trust model. The solution does not trust ISP for the communication content, while it trusts ISP for hiding the identity of users.

A: It is a problem of balancing. You have to leak some info to enable the communicate, while still you want protect the most important part which is the content that could be seen.

====

A NEaT Design for Reliable and Scalable Network Stacks 

Tomas Hruby (Speaker), Cristiano Giuffrida Lionel Sambuc Herbert Bos Andrew S. Tanenbaum 

There are two expectations for OS design: reliability and scalable.

The current commodity OS might be scalable but not reliable.

As the reliability is achieved by testing and bug fixing, but not originated from design scheme.

The speaker argues the the key for reliability is fault isolation.

However physical isolation is not performant as it introduces communication overhead.

The proposition of this work bases on microkernel, where system processes are spread over dedicated cores.

This is not a scaling design, as there could be bottle necks come from single component.

A nature thinking is to replicate those component to scale up, in a way that:

1/ application don't need to know the system replica numbers;

2/ replicas work independently;

3/ no sharing among replicas, so there is no locks.

The proposed network stack implementation allows the shared memory among network stack replicas.

Q: How network stack replicas work together, e.g. DHCP lease from different replica?

A: There is only one single ip for the entire stack. And communications took place among replicas.

Q: Comment on lock-free. 

A: It’s more about the way you programme, thread based or state-machine based.

Q: Fairness issue.

A: Components only run for a short time. Fairness can thus be achieved in a dynamic fashion.

====

ACACIA — Context-aware edge computing for continuous interactive applications over mobile networks

Junguk Cho (Speaker), Karthikeyan Sundaresan, Rajesh Mahindra, Jacobus Van der Merwe, Sampath Rangarajan

Continuous Interactive applications (CI) requires short latency.

Current enablers for CI:

1/ computation offloading to cloud;

2/ mobile networks, e.g. 5G;

3/ NFV & SDN;

4/ mobile edge computation;

5/ user context: location and interest of users.

Previous works mainly take a standalone approach.

The speak argues that it is not enough and thus proposes an end2end design, incorporating application, mobile network and user at the same time to delivery small delay for CI.

Then the speak presents the three components of the design.

1/ user context discovery

It uses publish and subscribe mechanism. 

User pushes interest to LTE modem filter.

The later then listens the broadcast and sees if there is a match.

2/ mobile edge network

In this part, the speak explains how CI apps establishes new connection to mobile edge network.

3/ context aware application

App aware of user location can reduce searching space and reduce latency.

LTE-direct publisher is used to locate user.

Then follows the implementation and evaluation.

Q: Localisation accuracy.

A: Not as accurate as dedicated approach; but the work has shown the feasibility and latency decrease already. 3 meter error; with more publisher, error will further decrease.

Q: How about combining WIFI with LTE-direct publisher?

A: The design is extensible. But the author don't want to include other radio technologies for power consumption considerations.

Q: There are two mobile cores: EPC + EPC in mobile edge. How the two cores synchronise for a single user?

A:  According to fig.5 in the article, the SGW-C & PGW-U insert rules into MEC basing on the user information to enable the connection from user to MEC.

=====

Lying your way to be better traffic engineering

Marco Chiesa (Speaker), Gábor Rétvári, Michael Schapira 

This work aims at enhancing the traditional TE scheme, in the situation where accurate knowledge on traffic demand is not available.

The problem is formulated as: given the network topology and traffic demand as input, construct per-destination DAG and traffic split that minimises the worst case link utilisation.

The speaker then decomposes the above problem into three steps:

1/ DAG construction.

In this step, shortest DAG is first calculated, then augmented with heuristics.

2/ traffic split in DAG

In order to reach optimality with reasonable computable complexity; the speaker casts the problem as a mixed linear-geometric programming problem.

3/ Legacy router configuration

One major contribution of the work is, when actual traffic deviates from the input traffic demand (i.e., with traffic uncertainty), the proposed scheme largely outperforms traditional TE scheme.

Q: Is split calculated for each TM snapshot over time?

A: The split is calculated for base TM.

Q: How it compares to source-destination networking?

A: They are two different settings. With source routing, you have more routing capability, but it is not compatible with traditional networking.

Q: What's the connection to SDN?

A: This work differs from traditional OSPF TE at these points 1/ explicit per dest DAG, instead of shortest DAG; 2/ per destination DAG and 3/ weighted in DAG split. Fibbing is just a way of implementing the proposed scheme. Other SDN configuration frameworks shall as well work. 

CoNEXT 2016: Session 6: Mobile systems and applications

The session chair Ganesh Ananthanarayanan started the proceedings by noting that this session will

officially be the longest session (albeit interesting) in the history of CoNEXT.

1st Paper -  Low Bandwidth Offload for Mobile AR

The first talk was from Puneeth Jain - a researcher in HP Labs. Puneeth mentioned that this work was his last as a PhD student, and thus this presentation "makes him feel like he is graduating today !!!".

Environmental fingerprinting has been proposed as a key enabler to Augmented Reality (AR). Although the environment fingerprinting can be done in many ways (e.g., wireless RF, magnetic field), visual fingerprinting is the most attractive option due to the inherent heterogeneity in many indoor spaces. However,  matching a unique visual signature against a database of millions requires either impractical computation for a mobile device, or to upload large quantities of visual data for cloud offload.

The authors propose a system called 'VisualPrint', which provides means to offload only the most distinctive visual data, that is, only those visual signatures which stand a good chance to yield a unique match. VisualPrint enables cloud-offloaded visual fingerprinting with efficacy comparable to using whole images, but with an order reduction in network transfer.

Questions Asked:

a) Have u considered applying for deep learning ?

    No, deep learning was not consierd. For deep learning one needs a huge dataset. Here, the goal was to create a visualprint using one image. The idea behind the paper was to create a social networking app where users could leave comments when they  see an image. With this being the purpose, the authors think their method is best suited.

b) Some of these AR apps are outdoors, when there are lot of people, specaially in a crowd. how does feature extraction happen.

    This issue has been handled in other work of their's. Not used in this work.

========================================================================

2nd paper -

D-Watch: Embracing “bad” Multipaths for Device-Free Localization with COTS RFID Devices

Speaker: Ju Wang

With more and more applications dependent on localization, the authors of this paper have come up with an interesting idea of device-free localization, to help applications like intrusion detection, elderly monitoring. This paper introduces D-Watch, a device-free system built on top of low cost commodity-off-the-shelf (COTS) RFID hardware. D-Watch leverages the “bad” multipaths to provide a decimeterlevel localization accuracy without offline training. In order to do this, D-watch harnesses the angle-of-arrival (AoA) information from the RFID tags’ backscatter signals.

Questions Asked:

a) How frequently can locations be attained? specially for a moving target?

        People generally move 1m to 2m per sec. The system gets data more frequently, and so can track people more accurately.

b) Can your system work in different enviornments (eg., if there are open spaces ? ).

         Yes, the authors expect their systems to work.

c) How much does people moving around impact accuracy? Did you consider caliberating the system when people are moving around.

            For initial caliberation, people are required to be static.

========================================================================

3rd paper -

Title: RT-OPEX: Flexible Scheduling for Cloud-RAN Processing

Speaker: Kassem Fawaz

With the Cloud-Radio Access Network (C-RAN), becoming the go-to architecture for deploying cellular networks in recent times, the paper discusses the challenges involved in designing baseband processing on these platforms. Since the signal processing is now implemented in the cloud, it becoomes imperative that the baseband signals are processed with minimal latency so as to ensure  deadlines of processing wireless frames are met., e.g., 3ms to transport, decode and respond to an LTE uplink frame.

The authors through their in-depth analysis show that commonly used (e.g., partitioned) scheduling techniques for wireless frame processing are inefficient as they either over-provision resources or suffer from deadline misses.This inefficiency stems from the large variations in processing times due to fluctuations in wireless traffic. The authors present a new framework called RTOPEX, that leverages these variations and proposes a flexible approach for scheduling. Evaluation of RT-OPEX is done on a commodity GPP platform using realistic cellular workload traces. Results show that RT-OPEX achieves an order-of-magnitude improvement over existing C-RAN schedulers in meeting frame processing deadlines.

Questions Asked:

a) What factors cause the  variation of processing time?

      channel time and mcs used affects the processing time.

b) Is the system sclable with increase in the number of UEs and number of enodB's?

      haven't verified the scalability, but the author doesn't expect the performance to drop when scaled.

========================================================================

4th talk:

Title of the paper - Enabling Automatic Protocol Behavior Analysis for Android Applications

speaker - Jeongmin Kim

Understanding app behaviour is crucial for network operations. However, this requires application layer protocol analysis which makes the task challenging. This paper presents Extractocol, the first system to offer an automatic and comprehensive analysis of application protocol behaviors for Android applications. Extractocol uses the android app binary as input and accurately reconstructs the HTTP request-response pair transactions.

Questions:

Since the speaker was unable to understand any of the questions asked, the eventual questions/discussions was taken offline.

CoNEXT 2016 - Session 10: Wireless 3

Title:
CPRecycle: Recycling Cyclic Prefix for Versatile Interference Mitigation in OFDM based Wireless Systems

Authors:
Saravana Rathinakumar (The University of Edinburgh)
Bozidar Radunovic (Microsoft Research UK)
Mahesh K. Marina (The University of Edinburgh)

Nowadays, OFDM is the most used data encoding method for wireless communications. The channel is split in sub-carriers, which are used in parallel, and a different sequence of symbols is transmitted over each sub-carrier. In order to avoid problems in recognizing subsequent symbols due to multipath effects, each symbol is preceded by a portion of its own tail, which is called cyclic prefix. Nevertheless, the length of the cyclic prefix is over-provisioned, resulting in a wastage of communication capacity.

The common decoding procedure is to discard the cyclic prefix and apply the FFT at the beginning of each symbol. Instead, the authors leverage the cyclic prefix in order to detect the best point in which the FFT should be applied. The key observation is that the FFT can be applied in every point between the beginning of the cyclic prefix and the beginning of the symbol. Indeed, the symbol decoded is the same, but the interference in the final result varies a lot between the different starting points.

The final idea is simple: detect the best starting point for each decoding, i.e. the one which results in less interference in the decoded output. This is done through two steps. First, a interference model is created using the preables that are transmitted for channel estimation at the beginning of the communication. Then, the model is used to estimate the starting point that maximizes the likelihood of the decoded symbol, i.e. the starting point whose output is less likely to be affected by interference. As the outcome of the technique is an improvement in the decoding process, it can be used for two purposes: reduction of adjacent channel interference, and reduction of co-channel interference.

The authors have tested the technique both through simulation and experimentation. The results are promising: the technique allows a reduction of up to 25 dB of the adjacent channel interference and up to 15 dB of the co-channel interference. Most important, the technique does not require any modification on the sender side, allowing an easy incremental deployment.

Q: What if the interference starts after the preables transmitted initially? Since the interference model is based on them, this could be a problem.
A: It is unlikely that the interference is not affecting at all the preables used in channel estimation.

Q: Why the interference model is based only on the initial preambles?
A: The authors want to create the interference model first, and then use it on the actual data reception.

Q: The technique seems to rely on the timeliness of symbol transmission, while what happens in reality is that each symbol might be transmitted a bit sooner or a bit later. How this affects the results of the technique?
A: This kind of time differences are taken into account to some extent by the interference model.


Title:
Leveraging Electromagnetic Polarization in a Two-Antenna Whiteboard in the Air

Authors:
Longfei Shangguan (Princeton University)
Kyle Jamieson (Princeton University & University College London)

In the field of human-machine interaction, in-air writing is an innovative proposal. The most recent solutions that realize in-air writing are based on passive RFID tags that are remotely tracked. In order to perform tracking, a trade-off has to be found between localization accuracy and costs: single-antenna solutions have high uncertanity in the output, while a higher number of antennas comes with higher precision and higher cost.

Instead of trying to individuate each exact position of the RFID tag, the authors propose to focus on getting the trajectory and the displacement of each movement. This approach allows to reduce the number of needed antennas to two. For trajectory detection, the different polarization mismatch perceived at the two antennas is used. For displacement computation, the authors use the triangular inequality to estimate a feasibility region, i.e. points in which the tag could have moved, and then they combine it with the direction previously computed to estimate the true displacement.

The authors have tested the approach both on whiteboard and in-air. The first approach has the advantage of "projecting" a 3d movement on a 2d plane, i.e. the influences in the movement given by the third dimension are avoided. On whiteboard, the technique allows to correctly recognize 15/26 letters with probability higher than 0.9, and exhibits an average recognition accuracy of 93.6%. Instead, the average recognition accuracy of in-air writing is a bit lower (around 83%).

Q: Could this technique be combined with other localization techniques?
A: The technique might be combined with other approaches, but its key advantage is that it does not require expensive antenna arrays. Instead, this kind of equipment is actually needed by the other solutions for localization.

Q: In the demo shown, the movements performed during tests are extremely slow. How does movement speed affect the results of the technique?
A: Currently, the system requires the movements to be slow in order to provide a good accuracy.

Q: Why some letters are not correctly recognized?
A: These letters (as C and L) are difficult to distinguish by the system as they have less evident differences.

Q: A possible extension to the work is the use of Wi-Fi to improve trajectory detection. Nevertheless, the polarization mismatches on which the current system is based are more likely to appear in short-range communications rather than long-range ones, as Wi-Fi is. How is the solution supposed to deal with this issue?
A: The authors confirms that the Wi-Fi extension might be problematic on this side.


Title:
Maximizing Broadcast Throughput Under Ultra-Low-Power Constraints

Authors:
Tingjun Chen (Columbia University)
Javad Ghaderi (Columbia University)
Dan Rubenstein (Columbia University)
Gil Zussman (Columbia University)

Networks of sensors which are able to harvest energy from the surrounding environment are becoming popular and popular. The sensors are required to communicate with other sensors as well, nevertheless the energy needed to transmit data is usually much more than the energy that can actually be harvested. In order to extend the life of the sensors as much as possible, severe constraints on the transmission and reception of data have to be imposed.

The authors consider a scenario in which sensors are heterogeneous, i.e. they have different energy budgets and different energy consumptions, and have no previous knowledge of other sensors nearby. Then, the authors model the scenario as a linear programming mathematical problem, whose aim is to maximize the global throughput. Finally, the authors use the results to design a protocol that leverages on broadcast communication and alternation between transmission, listening, and idle states in order to maximize the throughput, meeting the requirements dictated by the low-power constraints at the same time.

The results of the simulations show that the parameter sigma determines the trade-off between throughput and discovery latency in communication (the more sigma is high, the lower is the throughput, but the lower is also the discovery latency). The protocols allows to reach between 67% and 81% of the analytical maximum throughput, and performs 8x-11x better than the current state-of-the-art (i.e. Panda).

Q: In mobile ad-hoc networks, an approach is to keep a backbone of sensors always active and have the rest of the network mostly sleeping. Can this approach be compared to the presented one?
A: The presented approach do not require any gateway node to communicate, and this is an appreciable feature.

Q: The functioning of the protocol is based on the estimation of the number of listeners for each transmission. How is this estimated obtained?
A: After each transmission, a sensor puts itself in a short listening period to receive some kind of acknowledgement from the listeners. It can therefore use this number for the protocol.

Q: Low-power unicast solutions do exist. Which is the advantage of the broadcast transmission in this context?
A: Unicast solutions require the nodes to create and keep updated a routing table, which imply a non-negligible additional energy amount.

Q: The scenario analyzed assumes that nodes can always transit between transmission, listening and sleeping states. What if we consider a scenario where we have only-receiving nodes and only-transmitting nodes?
A: The protocol aims to coordinate the nodes such that a sensor will not wake up and put itself in listening if no data is available, nor it will put itself in transmission state if nothing has to be transmitted. Moreover, once the energy of a node goes down, the probability to transit to a transmission stage significantly drops, i.e. the node is "locked" in listening or sleeping state to the purpose of preserving energy.

CoNEXT 2016 - Session 7: Datacenter 2

Paper: Sunflow: Efficient Optical Circuit Scheduling for Coflows

Authors: Xin Sunny Huang (Rice University), Xiaoye Steven Sun (Rice University), and T. S. Eugene Ng (Rice University)

In one hand, Optical Circuit Switching (OCS) has many advantages, such as energy efficiency, cost efficiency, and its future-proof capabilities. On the other hand, it has worst traffic performance, especially for small data. This worst performance is related to the need to set up a circuit. In scenarios with large data, its performance may become closer to the one achieve by Packet Switching (PS). The paper seeks to answer the following question: can OCS be as good as packet switching for coflows?

The proposal of Sunflow is to use a not-all-stop switch model (instead of an all-stop model) to allow a more flexible scheduling of coflows. The proposed approach uses a greedy heuristic approach and it is within 2x the optimal. In practice, it is within 1.03x to the optimal. Regarding scheduling of intra-coflows Sunflow not allows subflows to preempt each other, while for inter-coflows it has a flexible preemption policy.

The evaluation was performed through simulations using traces from a Facebook cluster. The results showed that Sunflow is more efficient than Solstice and can achieve near packet switch performance when compared to Varys. In summary, Sunflow achieves the benefits of OCS and good traffic performance for coflows.

Questions:

Q: Is the 2x to the optimal for intra-coflows? What is your intuition about the worst case?

A: Yes. For small coflows performance becomes smaller. but the differences when compared to PS still small.

Q: Evaluation was made through simulations. What are your thoughts about deploying and evaluating in real scenarios?

A: There are other papers that address this question. Our contribution is on the algorithm.

Q: What is your opinion about the drawbacks of OCS?

A: Since OCS achieves many more additional benefits, the drawbacks are tolerable.

________________________________________________________________________

Paper: ECN or Delay: Lessons Learnt from Analysis of DCQCN and TIMELY

Authors: Yibo Zhu (Microsoft Research), Monia Ghobadi (Microsoft Research), Vishal Misra (Columbia University), and Jitendra Padhye (Microsoft Research)

Datacenter applications demand high bandwidth, low latency, and low CPU overhead. The TCP stack, however, is too heavyweight. RDMA have being employed in the real world to offload the NIC. Recently, DCQCN and TIMELY were proposed to improve the congestion control. This paper aims to answer the following question: which is better for congestion control?

While DCQCN achieves the desired properties (fairness, fast convergence, stability, high link utilization, and reduction in flow completion time), TIMELY does not. The reason is related to the use of the derivative of latency to change the rate. To deal with this situation, this paper proposes a quick patch for TIMELY. The patched version of TIMELY changes the rates based on absolute delay and on the derivative of delay.

Comparing the three solutions, DCQCN obtained better results than both versions of TIMELY. The patched version of TIMELY outperformed the original one. As the workload increases, the results of DCQCN become even better than the other approaches. The main reason for the performance difference is related to the mechanism used to detect the congestion. While DCQCN uses ECN, TIMELY is based on end-to-end delays. The conclusion is that the use of ECN is better because on delay-based approaches is not possible to have fixed queue length and fairness simultaneously. In summary, ECN is probably a better signal than end-to-end delay. However, the use of end-to-end delay does not need support by the switches.

Questions:

Q: Have you evaluated the fairness in scenarios where there are a variety of end-to-end delays?

A: No. The focus of the work is on datacenter, where the queueing delay is dominant.

Q: Have you considered the existence of multiple bottlenecks per flow?

A: Not yet.

Q: Have you tested with different flow sizes?

A: Not yet.

________________________________________________________________________

Paper: Composite-Path Switching

Authors: Shay Vargaftik (Technion), Katherine Barabash (IBM Research), Yaniv Ben-Itzhak (IBM Research), Ofer Biran (IBM Research), Isaac Keslassy (Technion), Dean Lorenz, (IBM Research), and Ariel Orda (Technion)

Datacenters are requiring lower latencies and higher bandwidths. How is it possible to achieve these demands? First, it is necessary to understand the current scenario. Today, we have lots of racks with small demands and several racks with lots of traffic. DCN traffic patterns can be classified in many-to-many, one-to-one, one-to-many, and many-to-one. The challenge is how to deal with the last two. This paper proposes a Hybrid Switching model, which combines EPS and OCS.

In order to avoid performance degradation, the one-to-many traffic should have OCS for the sender and EPS for the receivers. In the many-to-one case, the composition should be EPS for the senders and OCS for the receiver. Two main challenges should be tackled: how to represent composite paths and what to serve using composite paths. The first one is treated with an augmented demand matrix. The second one is achieved through the recognition of the different paths and its properly scheduling. The scheduling works in steps. First, a reduction process is applied on the demands and on the switch parameters. Second, the new demands and the hybrid-switching parameters are passed to the scheduler. Finally, the interpretation process schedules the flows. In summary, we proposed an approach that can accommodate more traffic patterns without increasing the scheduling complexity.

Questions

Q: It appears that the OCS and EPS will share the same path. Thus, the OCS capacity will be limited by the EPS. What is your position?

A: I do not see this limitation. There are extra high bandwidth ports to avoid this.

Q: How recently are you traces?

A: They are of 2014, maybe 2015.

Q: Does the larger piece of traffic originated by many-to-many communication? Is it possible to use the composed path for this traffic?

A: Yes. But we do not want to overhead the composed path.