Advisers: Myra B. Cohen and Sebastian Elbaum

To achieve "practical tractability," this thesis proposes: (1) New local consistency properties and algorithms for enforcing them without adding constraints or altering the network's topology; (2) Methods to enforce these consistency properties on the clusters of a tree decomposition of the CSP; and (3) Schemes to bolster the propagation between the clusters of the tree decomposition.

Our empirical evaluation shows that our techniques allow us to achieve practical tractability for a wide range of problems, and that they are both applicable (i.e., require acceptable time and space) and useful (i.e., outperform other consistency properties). We theoretically characterize the proposed consistency properties and empirically evaluate our techniques on benchmark problems. Our techniques for higher level consistency exhibit their best performances on difficult benchmark problems. They solve a larger number of difficult problem instances than algorithms enforcing weaker consistency properties, and moreover they solve them in an almost backtrack-free manner.

Adviser: Berthe Y. Choueiry

Adviser: Ziguo Zhong

To evaluate the algorithms, we simulate node failure recovery in the clusters of different sizes, construct the energy model and analyze the energy consumed during the failure recovery process. The simulation results show that the erasure coding based failure recovery algorithm often outperforms the replication based approach. On average, the former requires 60% of the energy as that of the later and the energy saving increases with the cluster size. In addition, with our recovery aware data placement strategy, the energy consumption for both approaches could be further reduced.

Advisor: Ying Lu

This thesis investigates the topic of user and data modeling for recommendation systems in a wiki environment. In addition to conventional usage data, the proposed model uses new metrics designed for the wiki domain, including active-passive activity level rating and minimalist-overachiever score. The active-passive activity level rating provides a quick overview of a participant’s collaborative activity composition and can be leveraged to alert moderators when participants aren’t meeting expectations. The minimalist-overachiever score strongly correlates to evaluations that participants have received and can be used as an aid in determining performance in future collaborations. These, along with other findings, serve as the foundation for improved virtual collaboration.

Adviser: Leen-Kiat Soh

In this thesis, we define a new Test Case Language (TCL) that can be used to represent test cases that vary in structure and are generated by multiple test generation frameworks. We also present a methodology for transforming test cases of varying representations into a common format where they can be matched and analyzed. With the common representation in our test case description language, we define five advice functions to leverage the testing strength from one type of tests to improve the effectiveness of other type(s) of tests. These advice functions analyze test input values, method call sequences, or test oracles of one source test suite to derive advice, and utilize the advice to amplify the effectiveness of an original test suite. Our assessment shows that the amplified test suite derived from the advice functions has improved values in terms of code coverage and mutant kill score compared to the original test suite before the advice functions applied.

Advisors: Stephen D. Scott and Jitender Singh Deogun

As our first step, we investigate the main sources of race detection overhead and find that a large effort is spent on repeatedly monitoring operations that cannot cause data races or have already been identified as causes of races. Based on these observations, we propose two orthogonal optimizations for race detection: Stationary Object Suppression (SOS) and Loop Iteration Sampling (LIS). SOS employs a dynamic program analysis technique to filter out Stationary Objects; which are read-only objects that can be shared by multiple threads. As such, they can never participate in data races. By eliminating monitoring operations on Stationary Objects, SOS can detect up to six times more races within an overhead budget than Pacer, a state-of-the-art sampling based race detector.

Although SOS can greatly reduce the number of objects to monitor for race detection, it repeatedly monitors identified sources of races. A further investigation shows that loops in a program substantially contribute to occurrence of such repetitive data races. We propose a sampling based race detector, LIS, which adjusts sampling rate for data access operations within loops to be inversely proportional to number of iterations.

To achieve perpetual availability, the next step is to address these software faults as they are detected during deployment. We propose a race healing system that can automatically generate and apply repairs during program execution. The system applies a fix immediately after a race is detected to prevent the race from occurring again.

Advisers: Witawas Srisa-an and Matthew B. Dwyer

Adviser: Lisong Xu

The idea of combining information computed by different analyses has been circulating in the research community since the mid 1960’s and has shown the benefits of analyses uni- fication. Several approaches have been developed for combining multiple static analyses, and combining static and dynamic analyses. These approaches mainly deal with combining the intermediate result of one analysis to help another analysis with deciding P |= φ. This dissertation takes an alternative approach by allowing each analysis to determine P |= φ under some conditions. Then, combining the final results of such analyses causes P |= φ to hold under a weaker condition until, ultimately, an unconditional final result is produced.

This dissertation formalizes and implements a unification framework that combines computed information from analyses and disseminates that information among other anal- yses. This framework is extensible since the only requirement that an analysis shouldsatisfy is to have querying and reporting capabilities. Conducted in this context of the uni- fication framework, our experiments have shown that combining results from a diverse set of analyses produces weaker conditions for P |= φ than analyses can achieve operating in isolation.

Adviser: Matthew B. Dwyer

For the private LLC organization, prior-art proposals can improve the efficacy of inter-core cooperative caching at the coarse-grained application level. However, they are still suboptimal because they are unable to take advantage of the diverse capacity demands at the fine-grained set level. We introduce the SNUG LLC design that exploits the set-level non-uniformity of capacity demands and thus further improves performance.

Still for the private LLC management, we notice that neither spatial nor temporal LLC management schemes, working separately as in prior work, can deliver robust performance under various circumstances due to set-level non-uniform capacity demands. We propose a novel adaptive scheme, called STEM, to solve the problem by interactively managing both spatial and temporal dimensions of capacity demands at the set level.

For the shared LLC organization, existing proposals try to improve either locality or utility for heterogeneous workloads. But we find that none of them can deliver consistently the best performance under a variety of workloads due to applications' diverse locality and utility features. To address the problem, we present the CLU LLC design that co-optimizes the locality & utility of co-scheduled threads and thus adapts to more diverse workloads than the prior-arts.

To make a cache management strategy practical for industry, we will need to cut the overhead of the re-reference prediction value (RRPV). We observe that delicately-tuned replacement policies rooted in single-bit RRPVs can closely approximate the performance of their correspondents with log{associativity}-bit RRPVs. Therefore, we propose a novel practical shared LLC design, called COOP, which entails a 1-bit RRPV per cacheline and a lightweight monitor per core for locality & utility co-optimization. At a considerably low storage cost, COOP achieves higher performance than the two recent practical replacement policies that rely on 2-bit RRPVs but are oriented towards locality optimization only.

Advisers: Hong Jiang and Sharad C. Seth

To address this problem, we introduce a new type of software properties called \textit{statistical properties}, which characterize significant statistical relationships among the values of variables across program states. We define an approach to infer these relationships automatically and support their monitoring while controlling the trade-offs between overhead and the precision and recall of the inferred properties.

We perform several experiments to assess the approach in the context of distributed robotics applications. Our findings indicate that the inferred statistical properties can be use to generate precise and cost-effective models capable of detecting faults in software systems while keeping the number of false positives close to zero and previous knowledge of the software system design and behavior unnecessary.

Adviser: Sebastian Elbaum

The first part focuses on the white-box approaches that optimize the internal design of SSD. We design a write buffer management algorithm on top of the log-block FTL, which not only optimizes the write buffer effect by exploiting both the recency and frequency of blocks in the write buffer, but also minimizes the destaging overhead by maximizing the number of valid pages of the destaged block. We further identify that the low garbage collection efficiency problem has a significantly negative impact to the performance of the page-mapped SSD. We design a GC-Aware RAM management algorithm that improves the GC efficiency even if the workloads do not have updating requests by dynamically evaluating the benefits of different destaging policies and adaptively adopting the best one. Moreover, this algorithm minimizes the address translation overhead by exploiting the interplay between the buffer component and the FTL component.

The second part focuses on the black-box approaches that optimize the SSD performance externally. As an increasing number of applications deploy SSD in enterprise environment, understanding the performance characteristics of SSD in enterprise environment is becoming critically important. We identify several performance anomalies of SSDs and their performance and endurance impacts on SSD employed in enterprise environment by evaluating several commercial SSDs. Our study provides insights and suggestions to both system developers and SSD vendors. Further, based on the performance anomalies identified, we design an IO scheduler that takes advantage of the SSD features and evaluate its performance on SSD. The scheduler is shown to improve performance in terms of bandwidth and average response time.

Adviser: Peter Revesz

Thin-plate spline warping is a non-linear technique that is widely used for registering different types of medical images including magnetic resonance and histology images. However, this technique is constrained by manual landmark selection. In this research, we have developed a method to automate the landmark selection process using thin-plate splines by maximizing the normalized mutual information between the two images. The approach has been studied in the context of registering MRI images and the histological sections of a rodent brain. The approach involves using level-set evolution to isolate the brain in a volumetric MRI image. Then the MRI volume is registered to the corresponding 3D histology (stacked histological sections) image using an affine transformation. The MRI volume is then re-sliced to match the corresponding histological sections. Finally these 2D MRI slices are warped to the histological sections using thin-plate splines maximizing the normalized mutual information. The approach was tested with images from 4 rodent brains with over 170 MRI images and over 120 block face images for each brain. The effectiveness of the landmark was determined by comparing its performance with the results manually obtained by three experienced technicians. The results show that the landmarks obtained using the NMI optimization approach is effective and comparable to manual extraction results.

Advisor: Ashok Samal

Adviser: Mehmet C. Vuran

In this paper, we propose an Indoor Positioning System (IPS) based on widely deployed indoor WiFi systems. Our system uses not only the Received Signal Strength (RSS) values measured at the current location but also the previous location information to determine the current location of a mobile user. We have conducted a large number of experiments in the Schorr Center of the University of Nebraska-Lincoln, and our experiment results show that our proposed system outperforms all other WiFi-based RSS IPSs in the comparison, and is 5% more accurate on average than others.

Advisors: Lisong Xu and Zhigang Shen

We provide a comprehensive survey that presents a picture of the current space networking technologies and architectures. In the survey, we discuss the IPN and Delay Tolerant Networking (DTN) concepts along with the various space networks that are currently deployed. We next propose a design of the IPN and implement it with the Interplanetary Overlay Network (ION) software module on real time physical nodes on the ORBIT testbed. Two space network scenarios are designed and experimentally evaluated to verify the correctness of the network implementation. We also focus on the study of bundle transmission delay and separately evaluate the effect of bundle size and number of bundles. The experimental evaluation provides insights into the factors which caused delay in bundle transmission such as custody refusal, expiration of bundle lifetime and congestion.

Adviser: Byrav Ramamurthy

First, we study the problem of heterogeneous user demands on video quality. We design an efficient P2P live streaming system which greatly reduces the bandwidth consumption and still achieves satisfactory streaming quality. To reduce the bandwidth consumption, we use adaptive streaming, so that the streaming rate of a peer is commensurate with its window size. In order to maintain satisfactory streaming quality even in the case when peers dynamically change their window sizes, small-window peers are designed to contribute part of their bandwidth to help large-window peers.

Second, we study the problem of heterogeneous user demands on playback delay. Specifically, we consider time-shifted streaming service and bounded-delay live streaming service. With time-shifted streaming service, a peer can watch a program which has been broadcast live before it joins the system. We propose a cooperation-based prefetching design which distributes video segments more widely in a P2P streaming system and thus greatly improves the streaming quality. With bounded-delay live streaming service, some peers (e.g., paid users) can watch the program with short and bounded playback delays, and other peers (e.g., free users) with best-effort short playback delays. We prove that the bounded-delay live streaming problem is NP-Complete, and propose a heuristic algorithm which constructs an efficient P2P overlay structure of peers by considering their playback delay requirements and their physical properties.

The promising results in this dissertation show that our algorithms are effective in meeting heterogeneous user demands on video quality and playback delay. In the future, we are interested in designing efficient P2P streaming systems in mobile and cloud environments.

Advisors: Lisong Xu and Byrav Ramamurthy

Advisers: Steve Goddard and Mehmet Can Vuran