Relevant Degree Programs
Affiliations to Research Centres, Institutes & Clusters
Graduate Student Supervision
Doctoral Student Supervision (Jan 2008 - May 2021)
As people increasingly use a wide variety of devices throughout their day, it becomes important to give users consistent and efficient access to their data in all their devices, while at the same time take into consideration the resource limitations in each device. This dissertation presents a storage platform that presents users with a single view of all their data, independently of the device they are using. Each of the user’s devices has full and up-to-date information about the entire data structure and metadata, and is able to retrieve any file transparently as needed. File content, including history, is efficiently distributed among the user’s own devices according to where it is used, with no need for centralized storage and no data stored in devices that are not owned by the user. Peers communicate with each other directly to maintain a consistent state. Users also have no need to know where the data is stored, or to remember in which device the latest update to a file has been performed. The user is able to adapt the platform to devices with different storage and network characteristics. While the latency to access content is affected, and some data may not be available at all times if some devices are disconnected, applications have access to tools that allow them to adapt to data availability.
In IEEE 802.11, nodes regulate access to the airspace they share in a decentralized fashion using CSMA/CA. The goal of this approach is to share the common airspace fairly and efficiently without requiring centralized channel administration or direct coordination among peer nodes. However, it is well known that strong interference, as consequence of this de-centralized coordination scheme, can lead to extremely unfair network bandwidth allocation between competing devices. Interference detection and mitigation has posed great challenges. The cause of interference is complicated, involving many networking factors such as topology and traffic, and the interference relationship changes all the time. This thesis addresses these challenges by proposing a throttling based interference mitigation system (Shaper) and an online passive interference detection system (VOID). The main contribution of this thesis is to point out the correlated relationship between interference and congestion. First, this thesis provides a more thorough analysis on the impact of node topology, traffic type and signal strength on wireless performance. We came up with 9 UDP models and 10 TCP models just for two competing flow scenarios. The outcome of wireless interference can get harder to predict, however, as we introduce more factors into the interference model such as more competing nodes, sending rate, signal propagation model, etc. On the other hand, this thesis identifies the immediate cause to the unfair bandwidth distribution under interference: 802.11 network congestion. We observed and explained that all competing devices are able to perform well regardless of topology or traffic class, as long as there is sufficiently more bandwidth than the aggregate throughput demands. Therefore, we propose to trade the aggregate throughput to mitigate the impact of interference and prove its effectiveness through simulation and emulation. Finally, the key to successful addressing the interference is an accurate and fast interference detection mechanism and this thesis proposes such a system called VOID. It deploys the correlation between congestion and interference to infer the interference relationship from the ip-layer throughput variations. It is fast, accurate and more importantly, very easy to deploy in existing WiFi networks.
This thesis describes the design, implementation, and evaluation of a dynamic program analysis framework called Tralfamadore. Three key aspects differentiate Tralfamadore from traditional dynamic analysis frameworks. First, analyses execute offline based on detailed CPU traces. This approach enables multiple analyses on the same execution, and analyses not foreseen at the time of execution. Using detailed traces also decouples analysis from instrumentation, simplifying the design of analyses and of the framework. Second, Tralfamadore supports the analysis of operating system execution. Third, the architecture of the analysis engine promotes the construction of analyses that are composable and reusable. New analyses can be written by mostly reusing existing components. Although these aspects have been investigated in isolation in the past, Tralfamadore is the first dynamic analysis framework to address all three at once.
Master's Student Supervision (2010 - 2020)
The "free-riders" issue happens in many peer-to-peer system, and this issue severely damages the liveness of the system especially for the system that heavily relies on coordination of multiple peers. Many peer-to-peer system assumes that the peers are willing to volunteer their resources to support functionality of the system, such as content storage and traffic relay. However, the assumption is hardly true in real Internet world, in which most of the users are trying to maximize their profit playing to take advantages from the system without providing corresponding services to others in the system. The notable decentralized anonymous network, TOR, also does not have any explicit incentives for the peer to host Onion Router as infrastructures relaying secure traffics to protect the other peers' anonymity. Therefore, eventually the behaving users lose their faith in the system due to the unfairness, and the system may cease to properly run.OnionCoin(OC) is the first decentralized peer-to-peer anonymous messaging system based on Onion Routing with the integration of currency system attempting to resolve such issues in Onion Routing based systems. A peer in OC will have to "pay" for using the anonymous messaging service, and will "gain" by helping other peers. The purpose of this currency system is to record the transactions of "pay" and "gain" in order to provide incentives for peers to participate more in the system. Several protocols are designed so that the introduction of the currency system does not compromise any existing protection in Onion Routing on the anonymity of the communicating parties.
Organising files by grouping them using tags is an alternative approach tofilesystem design that has benefits over the traditional hierarchical model.However, the majority of filesystems in use remain hierarchical. In this paperwe describe FindFS, an extension middleware which provides dynamic, tag-basedviews of an existing hierarchical filesystem. FindFS adopts the functionalityof the find utility and adds support for extended attribute queries, whichcan be used for tagging files and filtering the filesystem. Search results arerepresented as directories containing symbolic links which are kept up-to-datein response to filesystem operations, allowing them to persist as views thatare accessible from existing unmodified applications. Control of the system isaccomplished via filesystem operations, enhancing its ease of integration andportability. We have developed a prototype of this system, and characterisedthe performance overhead that it adds to file operations.
Parallax is a distributed storage system for virtualized environments. It provides virtual disk abstractions for the consumption of virtual machines. Parallax’s design has been driven by emerging trends in virtualized deployments. The system makes use of Copy-on-Write mechanisms to provide primitives for fast snapshotting and provisioning of disks in order to match the nimbleness and agility of virtual machines. However, using Copy-on-Write mechanisms in a distributed storage system complicates free space reclamation and leads to data fragmentation.Implementing a garbage collector and a disk defragmentation utility for Parallax has been attempted previously. The work described in this thesis presents a revised implementation of garbage collection for Parallax. It also introduces a utility for managing free space reclaimed from garbage collection, namely a fault-tolerant free space coalescing utility that aims at minimizing the occurrence of data fragmentation - and its adverse effects - in Parallax.
IEEE 802.11 wireless networks are widely deployed and used nowadays, especially in enterprise and university settings. The widespread usage means increased airspace congestion when many users connect at the same time, which slows down the wireless network performance. Two infamous problematic scenarios are the hidden terminal and the exposed terminal, where a victim node will suffer decreased throughput because of interference from other nodes which are unfairly using up the available airspace. This wireless unfairness occurs due to the design of the 802.11 protocol and the vast number of clients connected. The thesis presents a VOIDShaper Engine, which captures the network traffic at a central upstream router through which all wireless access points connect, and uses a previously developed VOID tool to analyze the traffic and generate a network interference map. With this map, VOIDShaper can find the interference between pairs of TCP Flows, and apply corrective traffic shaping rules on the interfering TCP flows directly at the router. All experiments were run in the Emulab testbed; we see that with traffic control, bandwidth fairness is achieved by using the HTB and SFQ queuing strategies at the router. The goals of this work are 1) to use VOIDShaper to detect and shape wireless traffic at the same 802.11 transmission rates due to interference, 2) to analyze how VOIDShaper responds to wireless interference when the nodes are transmitting at different wireless data rates, 3) to devise an algorithm and shaping policy for traffic which is sent at symmetric or different 802.11 transmission rates.