Improving resource efficiency of serverless with hierarchical isolation and resource-aware orchestration

While serverless computing provides a pay-per-use pricing model and enables developers to build scalable, event-driven applications, current platforms suffer from significant resource inefficiencies that increase costs and energy consumption. These inefficiencies stem from strong isolation mechanisms, limited resource sharing, slow startup paths, and conservative scaling practices that lead to persistent resource waste. This thesis explores architectural designs and system techniques to improve resource utilization without sacrificing security, performance, or compatibility. We investigate how WebAssembly (Wasm) can reduce isolation overhead while maintaining strong security guarantees. We introduce Wasabi, a system that incorporates a dense hierarchical architecture to securely co-locate Wasm applications from different customers within the same container sandbox. It preserves the container-based serving model, which allows leveraging existing platform capabilities and supporting non-Wasm-based workloads, reduces per-application overhead, smooths out fluctuations in demand across applications, and avoids long startup delays for most requests. Beyond the system itself, this thesis provides empirical insights into workload co-location, managing load spikes, and serverless capacity planning. This thesis also introduces several system techniques, including (1) resource-aware auto-scaler and admission control, (2) safe overcommitment, and (3) memory pre-allocation, optimizing resource efficiency while maintaining or even improving performance. Wasabi leverages these, complementing its hierarchical architecture and enabling higher deployment density. Beyond specific architectures, this thesis demonstrates how resource-aware orchestration improves serverless efficiency in container-based orchestrators, such as Knative. Additionally, it introduces a dynamic learning method for automated resource configuration, reducing common configuration errors. The proposed solutions contribute to more efficient and sustainable serverless, improving affordability and lowering energy consumption.