eCommons

 

Cornell Theses and Dissertations

Permanent URI for this collection

The theses and dissertations of graduate students at Cornell University have been deposited in Cornell's institutional repository (eCommons) since about 2004. This collection also includes a few earlier Cornell theses.

Students retain ownership of the copyright of their work. Students also have the option of imposing a temporary embargo on access to the full text of their theses for limited amount of time (see eCommons access policy). If access to a thesis is restricted, the metadata record for the thesis is still visible, but the text "Access to Document Restricted" is displayed, and a field labeled "No Access Until," which indicates the date when the full text of the thesis will become accessible.

More information about finding Cornell theses and dissertations is available on this library guide, and the eCommons help page for finding content in specific collections, including theses and dissertations.

In general, older theses and dissertations from Cornell University are not currently available as digital files in eCommons. The Library is willing to digitize and make available older Cornell theses on a cost recovery basis. If you are interested in this service, please contact dcaps@cornell.edu.

Browse

Recent Submissions

Now showing 1 - 10 of 10673
  • Item
    Vowel harmony and coarticulation in three dialects of Yorùbá: Phonetics determining phonology
    Przezdziecki, Marek A. (2005)
    This dissertation examines the phonology and acoustic phonetics of vowels in three dialects of Yoruba—Standard Yorùbá, Mòbà, and Àkùré Yorùbá—to investigate the role of coarticulation in the phonologization of vowel harmony (Ohala 1994). The phonological vowel patterns of the three dialects are presented. Àkùré Yorùbá exhibits Advanced Tongue Root (ATR) vowel harmony in mid and high vowels, while harmony in Mòbà and Standard Yorùbá does not extend to high vowels. In order to investigate this relationship, recordings of VCV nonsense words from speakers of each dialect were analyzed. Following Hess (1992), the first formant (F1) was determined to be the acoustic measurement best correlated to the ±ATR vowel sets. Other measurements—F2, F1 bandwidth, fundamental frequency, vowel duration, and spectral measures—were not found to correlate with ATR. Using F1 as a measure, vowel to vowel coarticulation in high vowels in Mòbà and Standard Yorùbá was found to resemble high vowel harmony in Àkùré in the target vowels, the context, and the phonetic effect. This was particularly true for /i/; however the coarticulatory effects on /u/ were weaker and not statistically significant. As expected, the effect of vowel to vowel coarticulation in Mòbà and Standard Yorùbá was smaller and less robust than for vowel harmony in Àkùré. A decision tree model is proposed that is able to generate the high vowel harmony pattern from the Àkùré acoustic data. More interestingly, the model succeeds at extracting—to a large degree—the high vowel harmony pattern from Mòbà and Standard Yorùbá, the dialects without high vowel harmony. The model does not require any reference to features or natural classes, suggesting that it is not necessary to posit features as a prerequisite to learning a phonological pattern, nor as an explanation for universal patterns. The study argues that the acoustic patterns found in vowel to vowel coarticulation are sufficient to result in vowel harmony. The findings are consistent with the view that proto-Yorùbá did not have harmony in its high vowels (Fresco 1970, Oyelaran 1973, and Capo 1985), and that high vowel harmony developed in Àkùré and related dialects.
  • Item
    SYSTEMS ENGINEERING FOR NEXT GENERATION SPACE ENTERPRISES
    Zucherman, Aaron (2023-05)
    We are on the threshold of a new era of sustainable exploration and development of space. New Launch vehicles and programs such as NASA's Artemis and Lunar Gateway will change space technology and the stakes for space systems as we know them. As a result of these and other transformative changes, opportunities to launch and operate new space vehicle architectures will be unprecedented. To this end, this dissertation provides three foundational studies intended to impart rigor and systems thinking to the development and planning efforts of next-generation space projects. 1) Navigating the Policy Compliance Roadmap for Small SatellitesThis study explores USA space policy and regulatory processes and how they apply to satellites not fitting the typical mold of traditional missions. It lays out a systematic way forward for small satellite mission developers and managers to navigate the approval quagmire for individual spacecraft on multi-payload launches. It also puts forth ways for approving new and expected future mission architectures and technologies. Additionally, areas are identified where there are policy and regulation gaps and "gray areas" to prepare developers and inform other stakeholders of potential issues. 2) Lessons Learned from the First Generation Interplanetary CubeSatsThis study analyzes information gathered from the first sixteen interplanetary CubeSats and the unique difficulties faced by this mission type. Solutions to the specific development problems and general observations on the engineering and programmatic challenges faced by this mission type were solicited from previous mission developers and documented. From this, development approaches are proposed to lower risk and costs for future mission developers and stakeholders. 3) Evaluating Mars Rotorcraft Development InvestmentsRotorcraft can offer a new paradigm for Martian surface and atmospheric exploration missions. This study was conducted to enable stakeholders to evaluate competing research and development efforts for Mars Rotorcraft technologies. Not only for their estimated costs and system performance but also for their long-term improvement potential in the context of other ongoing developments. It does this by establishing critical metrics and relationship models for evaluating rotorcraft system and subsystem performance. Then the alignment of potential developments to the broader NASA technology goals and ways to estimate returns on investments were established.
  • Item
    MODEL-BASED CONSTRAINTS ON NUTRIENT CYCLING IN THE GLOBAL ENVIRONMENT
    Zhou, Yanqiu (2023-05)
    This dissertation has three chapters. Chapter 1 examines nutrient resupply patterns during decomposition in forest ecosystems, including tropical, temperate, and boreal, through meta-analysis. The hypothesis tested is that C, N, and P follow different prototypes in mineralization and be affected by the mean annual temperature and precipitation of sites. Results show that P will be preferentially released compared with C in mineralization, while C and N are coupled and released together. And C is more obviously affected by the higher the mean annual temperature (MAT), the higher the mineralization rate. C shows a significant increase in the mineralization rate with increasing temperatures. At the same time, N and P are not as strong as C. Thus, global climate change will aggravate the loss of C, further worsening the greenhouse effect. However, mean annual precipitation (MAP) has no significant effects on it. Chapter 2 analyzes the nutrient (N and P) use efficiency, global fertilizer uses for 2015, and predictions for the year 2050 using models and scenario analysis. Country-level nutrient use efficiency was calculated based on crop yield and total nutrient inputs for each country, and global heterogeneity was studied. Five scenarios were applied for 2050 fertilizer demand prediction: business as usual (BAU), climate change mitigation, nutrient use efficiency improvement, dietary shift, and all methods. Results showed that some countries in Africa and South America have abnormally high nutrient use efficiency, which indicates nutrient mining. Generally, nutrient use efficiency is higher in developed countries and lower in developing countries. For fertilizer use, by the year 2050, even population grows over 30 percent, with all scenarios applied, the fertilizer use can still reduce while feeding the population. Chapter 3 studies technology and management that can increase the nutrient (N and P)use efficiency, and did a meta-analysis and scenario analysis. Meta-analysis results were applied as nutrients use efficiency increasing scenario to fertilizer application in the year 2050. The results show that technologies and management can reduce future fertilizer demand. If combined with the scenarios in Chapter 2, the fertilizer demand in 2050 can be even less than in 2015.
  • Item
    The Surprise Factor: A Semantic Theory of Mirativity
    Zhuang, Lingzi (2023-05)
    This dissertation is a systematic study of the semantics of mirativity, a descriptive category for surprise-related meanings and their expression in natural language. Building on both typological and formal research traditions, I argue for a concerted research program for investigating this domain of meaning. Specifically, I address three foundational questions: (i) what is mirativity (semantic identity); (ii) what is the content of mirative meaning, and (iii) how does mirative meaning arise from evidential meaning.Despite a strong intuition that linguistic expressions of surprise form a natural category (mirativity), existing literature has persistently faced a certain amount of ambiguity over the definition and usage of this notion. I first resolve this ambiguity by articulating a semantic definition of mirativity grounded in the cognitive science of SURPRISE: mirativity is a range of attitudes which characterize mental states induced by the experience of SURPRISE. These atti- tudes necessarily contain a dimension of either novelty or counterexpectation: the latter causally induces SURPRISE, and the former is a necessary condition of the latter. Second, previous work has shown mirative meaning to have either propositional or speech act-level content. I argue that there is a third typological possibility: novel data on the mirative marker yikaon in Shanghai Wu (Sinitic, China) show that the content of a mirative attitude can be the union of a set of propositions: such miratives can crucially predicate an attitude over both single propositions in the declarative and questions with non-trivial informative content. I analyze this mirative contribution as an emotive attitude update to the speaker’s Discourse Commitments, which scopes over sentential force. Third, across languages, mirative markers are often also evidentials. I argue that the semantic affinity between evidentiality and mirativity has diverse theoretical characters: indirect, reportative and inferential evidentials do not evoke mirative meaning in the same way. Specifically, I argue that the connection between reportative evidentiality and mirativity can be due to diachronic reanalysis. Reportatives frequently trigger Conversational Implicatures about the SPKR’s attitude because they encode perspectivally asymmetric Discourse Commitments and QUD-addressing proposals. I argue that reportatives are often reanalyzed as SPKR-attitude markers due to the conventionalization of such Conversational Implicatures, driven by a principle of EPISTEMIC TRANSPARENCY.
  • Item
    ESSAYS ON ENVIRONMENTAL POLICIES IN THE TRANSPORTATION SECTOR
    Zhou, Hui (2023-05)
    This dissertation consists of three essays studying the effects of environmental regulations and policies in the transportation sector in China. The first chapter studies the effectiveness, efficiency, and distributional effect of using trade restrictions on used vehicles to protect the local environment. Leveraging comprehensive data on the bilateral trade of vehicles across Chinese prefecture cities and the staggered rollout of import restrictions on used vehicles implemented by city governments from 2013 to 2015, this chapter shows empirical evidence that import restrictions reduce net imports of used vehicles, and cities’ import restrictions are strategic complements. With a multi-sector multi-region structural trade model, this study shows that unilaterally restricting imports of used vehicles leads to welfare trade-offs between economic losses vs. environmental benefits. Restricting heavy-polluting vehicles makes some cities better off, especially lower-income cities. However, decentralized restrictions are socially inefficient due to strategic interactions, and the effectiveness and efficiency of using import restrictions as an environmental instrument are limited compared to emission taxes. The second chapter, joint with Jie Bai, Danxia Xie, and Shanjun Li, explores the import restrictions on used vehicles in China from the perspective of local protectionism. Leveraging the universe of new and used vehicle registration/sales data and the staggered removal of the restriction across cities from 2016 to 2018, this analysis shows that the removal of restriction led to a sharp increase in the cross-city flow of used vehicles but had no significant impacts on local air quality in the short run. Interestingly, the new vehicle market points to a prisoner’s dilemma among city governments: a unilateral removal of the policy would reduce new vehicle sales in a city but increase new vehicle sales in other cities. The effect is stronger in cities with a large automobile industry. The findings highlight alternative motives behind local environmental regulations and the need for coordinated efforts at the national level. The third chapter, joint with Shanjun Li, Xianglei Zhu, Yiding Ma, and Fan Zhang, examines the effectiveness of various policy measures that underlie the rapid development of the EV market in China, based on detailed data on EV sales, local and central government incentive programs, and charging stations in 150 cities from 2015 to 2018. This research finds that consumer subsidies for vehicle purchases accounted for more than half of EV sales in China. Nevertheless, investments in charging infrastructure were much more cost-effective than consumer subsidies. An inexpensive policy that merely provided EVs with a distinctive, green license plate was strikingly effective. These findings demonstrate the varying efficacy of different policy instruments and highlight the critical role of the government in promoting clean technologies.
  • Item
    Radio Near-field Motion Sensing: Theory and Applications
    Zhou, Jianlin (2023-05)
    The COVID-19 pandemic brings us a critical reminder: our medical system is weak under the attack of a highly contagious disease, and medical resources are too limited to provide timely diagnosis and treatment for a large patient population. During the pandemic, the demand for home and wearable medical sensors, such as antigen sensors and finger pulse oximeters, had explosive growth. Furthermore, with the help of artificial intelligence, early detection and health status become feasible if sensors can provide continuous vital signs or bio-marker data. This is a potential indicator that with the maturing of digital medical sensors, hospital-centric disease diagnosis and prognosis can gradually migrate to home in daily life. Electromagnetic (EM) waves of different frequency bands have long been used in medical settings for imaging and diagnostic purposes, such as magnetic resonance imaging (MRI), Computer tomography (CT) scan, and microwave tomography. Advanced research in EM technology has been a key driver in pushing the frontier of medical and biological research in which the interaction of EM fields with biological systems has played an essential role. However, the high cost, system complexity, large size, and health hazards of radiation exposure make these state-of-the-art medical equipment not feasible for daily at-home uses and continuous health monitoring. In the past four decades, microwave sensing has been enhanced by significant hardware advancement and miniaturization. Researchers have applied various radio-frequency (RF) technologies, such as far-field radar and impedance tomography, in health monitoring and have significantly improved vital-sign monitoring in humans and animals. The change of target geometry and material property are represented by the backscattered radio wave features, such as the frequency, magnitude, phase, and radar cross-section (RCS). However, while enabling fully non-contact operation, most far-field radar sensors are monostatic and often require a direct line-of-sight (LOS) with the target\textquotesingle s moving surface. Furthermore, most of these far-field radar systems have large sizes and high power consumption, which limit their usage for mobile applications. Unlike far-field radar sensors, wearable RF sensors, whether by active units or passive tags, allow subjects to move freely during monitoring and couple more EM energy directly inside the body in the local near-field zone. This work focuses on this near-field RF sensing method to accurately measure physiological organ and tissue motion and contributes two essential aspects of design optimization and wearable applications: 1) In the near-field RF sensing theory, we first analyze the near-field radio and dielectric object interaction mechanism and propose a backscatter field model for near-field RF sensors, which considers the scattering field, permittivity, object size, sensor positions, and carrier frequency. Second, we analyze the composition of the Rx radio signal and explicitly explore how the time-invariant component affects the morphology of the complex signal's magnitude and phase representation. To optimize signal features and regulate the waveform morphology for consistent signal interpretation, we propose a complex vector injection algorithm. 2) In the near-field RF sensing application, we introduce the design of low-power wearable near-field RF sensor platforms and their applications on human and animal vital signs sensing.
  • Item
    NEW PERSPECTIVES ON CONTINUOUS OPTIMIZATION: THEORY AND METHODOLOGY
    Zhao, Shipu (2023-05)
    Large-scale continuous optimization arises in many practical problems such as machine learning, signal processing, and imaging. It is usually challenging to analyze the theoretical properties of optimization algorithms and design scalable algorithms that work well in practice. This dissertation provides new perspectives on continuous optimization in both theory and methodology. From the theoretical side, we present a framework for reasoning about equivalence between a broad class of iterative convex continuous optimization algorithms. The notion of algorithm equivalence can make it easier to understand the connections between optimization algorithms, relate many analytical properties of interest such as convergence or robustness, and further give insights to design new algorithms. From the methodological side, we first present NysADMM, a scalable algorithm for faster composite convex optimization by exploiting the low-rank structure. NysADMM accelerates the inexact linearized Alternating Direction Method of Multipliers (ADMM) by constructing a preconditioner for the ADMM subproblem from a randomized low-rank Nystr{\"o}m approximation. Second, we present GeNI-ADMM, a framework generalized from NysADMM, which encompasses any ADMM algorithm that solves a first- or (generalized) second-order approximation to the ADMM subproblem and allows inexact subproblem solves. It facilitates the theoretical analysis of various approximate ADMM schemes for large-scale composite convex optimization.
  • Item
    RF SENSORS FOR MEDICAL AND CYBER-PHYSICAL INTELLIGENCE
    Zhang, Zijing (2023-05)
    My research has focused on continuous and non-invasive sensing of physiological signals including respiration, muscle activities, heartbeat dynamics, and other biological signals. I seek to establish a touchless RF sensor that can be implemented as wearables on users, or integrated into the furniture to become invisible to the user. Such sensor can greatly enhance data continuity, comfort and convenience to enable many healthcare applications, especially for at-home continuous diagnosis and prognosis, with less reliance on subjective self report. My research utilized machine-learning (ML) algorithms that can take the physiological data from our sensors to provide holistic diagnostics and prognosis. This sensor has been applied to pulmonary diseases including COVID-19 and chronic obstructive pulmonary diseases (COPD) to help identify dyspneic exacerbation, leading to early intervention and possibly improving outcome. The sensor has also been applied to prevalent sleep disorders such as apnea and hypopnea. Another aspect of my research focuses on muscle monitoring. Conventional electromyography (EMG) measures the neural activity during muscle contraction, but lacks explicit quantification of the actual contraction. I proposed radiomyography (RMG), a novel muscle wearable sensor that can non-invasively and continuously capture muscle contraction in various superficial and deep layers. Continuous monitoring of individual skeletal muscle activities has significant medical and consumer applications, including detection of muscle fatigue and injury, diagnosis of neuromuscular disorders such as the Parkinson’s disease, assessment for physical training and rehabilitation, and human-computer interface (HCI) applications. I verified RMG experimentally on a forearm wearable sensor for extensive hand gesture recognition, which can be applied to various applications including assistive robotic control and user instructions. I also demonstrated a new radiooculogram (ROG) for non-invasive eye movement monitoring with eyes open or closed. ROG is promising for gaze tracking and study of sleep rapid eye movement (REM).
  • Item
    INTEGRATED ELECTRONICS ON SINGLE CRYSTAL ALUMINUM NITRIDE: LOGIC, MEMORY AND RF COMMUNICATION
    Zhang, Zexuan (2023-05)
    Although the ultra-wide bandgap (UWBG) semiconductor aluminum nitride (AlN) has been routinely used in optoelectronics such as deep ultraviolet (DUV) light emitting diodes (LEDs) and lasers, its adoption in RF and power electronics remains largely unexplored. This dissertation demonstrates the significant potential of AlN platform for the next generation electronics to merge logic, memory and RF communication, the three pillars of electronic information systems. Built upon structurally-pure single crystal AlN substrates with high thermal conductivity, the high and best-balanced performances of p-channel and n-channel devices presented in this dissertation challenge the common misconception that high performance complementary logic is impossible with nitride semiconductors. Together with the radio frequency (RF) filters that have already been demonstrated on the same platform thanks to the high piezoelectricity of AlN, the results in this dissertation are expected to enable a fully-integrated monolithic RF signal processing solution on AlN. In this dissertation, utilizing polarization engineering, conductive channels are generated on the electrically-insulating AlN by adding a thin layer of GaN or AlGaN on top. Taking advantage of the capability to maintain sharp heterointerfaces by molecular beam epitaxy (MBE), combined with the state-of-the-art fabrication process, devices with record performances were achieved. First, the observation and properties of polarization-induced 2d hole gases (2DHGs) in GaN/AlN heterostructures on metal-polar single crystal AlN substrates are presented. The reduced dislocation densities on single crystal AlN substrates compared to foreign substrates such as SiC and sapphire improve hole mobility, and a record high hole mobility of ∼ 280 cm2/V·s is measured at 10 K. By leveraging the highly-conductive 2DHG in conjunction with a highly-scaled 3D-gating architecture, the fastest (fT/fMAX = 25/45 GHz) p-channel FinHFETs that deliver record-high on-currents(ION =1.3A/mm at room temperature) and >0.5 W/mm at 6 GHz RF output power are achieved. This is the first time RF output power has been obtained in nitride pFETs, marking the entrance of nitride transistor technology into the new frontier of RF CMOS. Next, by flipping the polarity of the structure, 2d electron gases (2DEGs) can be induced in N-polar GaN/AlN heterostructures. Unlike metal-polar AlN, the homoepitaxy of N-polar AlN is challenging, primarily due to the high reactivity of N-polar AlN surface. A new atomic surface cleaning technique— Al-assisted surface cleaning is developed and enables MBE homoepitaxy of electronic- and optical-grade N-polar AlN. 2DEGs are successfully observed in N-polar GaN/AlGaN heterostructures with sheet resistances of ∼300 Ω/□. These are among the lowest sheet resistances reported in III-nitride 2DEGs. The first N-polar HEMTs on AlN are enabled by these 2DEGs, showing a high on-current of 2.6 A/mm, a high speed (fT/fMAX = 68/100 GHz) as well as >3 W/mm RF output power at 6 GHz. While there is large room for future improvement, these exciting performances already demonstrated in the first generation devices mark important milestones towards highly reliable RF electronics with excellent thermal management based on N-polar AlN HEMTs. Finally, as an important first step towards direct integration of magnetic memory on the same nitride semiconductor platform, the MBE growth of ferrimagnetic Mn4N hosting desirable properties for spintronic applications is explored on GaN to form a all-nitride ferrimagnet/semiconductor heterostructure. Through exploration of nucleation and growth conditions, the MBE growth condition for c-axis aligned Mn4N on GaN with smooth surface morphologies is uncovered. Instead of direct nucleation of Mn4N on GaN substrates, a homoepitaxial GaN buffer layer is found to be helpful for improving the quality and surface morphology of the Mn4N epilayer. The ferrimagnetism evidenced by the clear anomalous Hall hysteresis loops at room temperature, along with the smooth surface morphology of Mn4N on GaN, lay the groundwork for bringing magnetic memory onto the AlN platform. As an aside, because of the rich magnetic phases within the Mn-N system and the large potential of antiferromagnetic materials for future spintronics, the magnetic properties of antiferromagnetic MnN are studied using optical second harmonic generation (SHG). The point group symmetry of MBE grown antiferromagnetic MnN films is identified as 2/m1’ and a loose upper bound on the domain sizes of 0.65 μm is placed. These results demonstrate the effectiveness of SHG for detecting the Neel order in metallic antiferromagnets and are expected to contribute to the recent efforts in using antiferromagnets for spintronic applications. With the results presented in this dissertation, the vision of fully-integrated electronics capable of logic, memory and RF communication functionalities on the single crystal AlN platform is very close to being realized, and will hopefully be greater than the sum of its parts.
  • Item
    EFFICIENT RESOURCE MANAGEMENT OF CLOUD NATIVE SYSTEMS
    Zhang, Yanqi (2023-05)
    Cloud native architecture has been a prevailing trend and is widely adopted by major online service providers including Netflix, Uber and WeChat. It enables applications to be structured as loosely-coupled distributed systems that can be developed and managed independently, and provide different programming models, namely microservice and serverless, to accommodate different user requirements. Specifically, microservices are a group of small services that collectively perform as a complete application. Each microservice implements a web server that handles specific business logic, and is usually packaged in a container that encapsulates its own runtime and dependencies. Microservice containers typically live for a long time and scale up or down to cope with load fluctuations as per user-specified policies. Serverless provides a further simplified approach to application development and deployment. It allows users to upload their application code as functions, without the need for explicit provisioning or management of containers, through an event-driven interface. Serverless containers are typically short-living ’one-off’ containers handling a single request at a time. The billing of serverless is fine-grained and users only pay for the resources consumed by actual function execution. Despite the popularity of cloud native systems, managing their resources efficiently is challenging. Cloud native applications consist of many component services with diverse resource requirements, posing a greater challenge compared to traditional monolithic applications. Furthermore, the backpressure effect caused by inter-service connections also complicates resource management. Lastly, although cloud-native relives users from the burden of infrastructure management, cloud providers still need to provision and pay for the infrastructure to host cloud native applications, which incurs high cost. This dissertation aims to tackle the challenge of efficient resource management for cloud-native systems and proposes three resource managers. First, we present \textbf{Sinan}, a machine learning (ML)-driven and service level agreement (SLA)-aware resource manager for microservices. Sinan uses a set of validated ML models to learn the per-service resource requirements , taking into account the effects of inter-service dependencies. Sinan's ML models predict the end-to-end latency of a given resource allocation, and the resource manager then chooses the optimal resource allocation that preserves the SLAs, based on the predictions. Sinan highlights the importance of a balanced training dataset that includes an equal share of SLA violations and satisfactions, for the effectiveness of ML models. Additionally, Sinan demonstrates that the system is flawed if the training dataset is dominated by either SLA satisfaction or violation. In order to obtain a balanced training dataset, Sinan explores different resource allocations with an algorithm inspired by multi-arm bandit (MAP). Although Sinan outperforms traditional approaches such as autoscaling, it requires a lengthy exploration process and triggers a large number of SLA violations, hindering its practicality. Furthermore, the ML models are on the critical path of resource management decisions, limiting the speed and scalability of the system. To address these limitations, we further propose \textbf{Ursa}, a lightweight and scalable resource management framework for microservices. By investigating the backpressure-free conditions, Ursa allocates resources within the space that each service can be considered independent for the puropose of resource allocation. Ursa then uses an analytical model that decomposes the end-to-end latency into per-service latency, and maps per-service latency to individually checkable resource allocation threshold. To speed up the exploration process, Ursa explores as many independent microservices as possible across different request paths, and swiftly stops exploration in case of SLA violations. Finally, in order to reduce the infrastructure provisioning cost of cloud-native systems, we propose to leverage harvested resources in datacenter, which cloud providers provide at a massive discount. Orthogonal to the first two parts of the thesis which aim to reduce operation cost by providing the minimum amount of resources that do not compromise performance, this part aims to achieve cost reduction by using cheaper but less reliable resources. We use serverless as the target workload, and propose to run serverless platforms on low-priority Harvest VMs that grow and shrink to harvest all the unallocated CPU cores in their host servers. We quantify the challenges of running serverless on harvest VMs by characterizing the serverless workloads and Harvest VMs in production. We propose a series of policies that uses a mix of Harvest and regular VMs with different tradeoffs between reliability and efficiency, and design a serverless load balancer that is aware of VM evictions and resource variations in Harvest VMs. Our results show that adopting harvested resources improves efficiency and reduces cost significantly, and request failure rate caused by Harvest VM evictions is marginal.