When I finally sat down to write this dissertation, the memories of the past six years unfolded before my eyes like scenes from a film. The countless debates over every idea, the tireless efforts behind each experimental result, the repeated rehearsals for every talk—all of it, suddenly, seemed to find its meaning.
In order to leave no regrets from these precious years, I poured my heart into perfecting every detail. Yet even after giving my all, I know that, as a researcher, I still have much to learn and to grow. I am deeply grateful to my two advisers for their understanding and patience, and for walking beside me and witnessing every step of this journey.
I hold deep admiration for my adviser, Sebastian Angel, whose unwavering pursuit of fundamental truths and boundless passion for new research ideas have been a constant source of inspiration to me. Without his clear and insightful guidance on structuring a paper, his thoughtful and timely feedback on my writing, and his unflagging enthusiasm during our countless discussions, I would not have grown into the researcher I am today.
I am also grateful to have had the opportunity to learn from and work with my former adviser, Joe Devietti. During times when my path in research felt dim and solitary, or when my research ideas were met with doubt, Joe’s steady support and unconditional trust once served as a guiding light, restoring my confidence and giving renewed purpose to my journey.
Heiner Litz, Tanvir Ahmed Khan, Saba Jamilan, Gilles Pokam, and Baris Kasikci — without the support of these external collaborators, the journey through a PhD would have been incredibly difficult to complete alone. I am deeply grateful to Heiner Litz, who took on the responsibility of mentoring me during Joe’s absence. His valuable insights on data prefetching, as well as his suggestion to integrate data prefetching features into BOLT for online prefetching, greatly shaped my thinking. I am also deeply thankful to Tanvir Ahmed Khan. When I struggled to naturally communicate my ideas to Joe at the beginning of my research journey, Tanvir served as a bridge between us, facilitating our discussions and easing communication. Throughout our collaboration, Tanvir contributed many bold and creative research ideas. His initial proposals of a tunable prefetch distance and rollback when prefetching proved harmful later became one of the core ideas behind RPG2. I am especially thankful to Saba Jamilan, who provided tremendous help in setting up the APT-GET system. I am grateful to Gilles Pokam and Baris Kasikci for their valuable feedback, which helped improve both OCOLOS and RPG2.
I would also like to thank my peers who supported me throughout this journey. I am grateful to Nathan Sobotka for working with me until the final hours leading up to the RPG2 paper submission deadline. I also thank Sehyeok Park, without whose help we would not have discovered MySQL, which became the most important benchmark for OCOLOS. I am further thankful to Haoran Zhang for his endless patience in teaching me how to debug containers of serverless functions, and for his valuable insights into selecting an orchestrator and a suitable serverless runtime for initiating our experiments for Quilt. I am grateful to Jess Woods for providing assistance with using linear programming solvers to resolve the merging decision problem in Quilt.
During the darkest time, Xinyi Chen, Megumu Tamura, Maggie Liu, Linda Chen, and Zhiyan Lu — with Megumu Tamura being my Japanese instructor and the latter three classmates from Penn’s Japanese Language Program — were there, listening to many of my complaints and encouraging me to stay positive. I am deeply grateful to the friends who stood by me during those difficult days.
Finally, I thank my parents for their unconditional support.