QoS-aware mechanisms for improving cost-efficiency of datacenters

Warehouse Scale Computers (WSCs) promise high cost-efficiency by amortizing power, cooling, and management overheads. WSCs today host a large variety of jobs with two broad performance requirements categories: latency-critical (LC) and best-effort (BE). Ideally, to fully utilize all hardware resources, WSC operators can simply fill all the nodes with computing jobs. Unfortunately, because colocated jobs contend for shared resources, systems with high loads often experience performance degradation, which negatively impacts the Quality of Service (QoS) for LC jobs. In fact, service providers usually over-provision resources to avoid any interference with LC jobs, leading to significant resource inefficiencies. In this dissertation, I explore opportunities across different system-abstraction layers to improve the cost-efficiency of dataceters by increasing resource utilization of WSCs with little or no impact on the performance of LC jobs. The dissertation has three main components. First, I explore opportunities to improve the throughput of multicore systems by reducing the performance variation of LC jobs. The main insight is that by reshaping the latency distribution curve, performance headroom of LC jobs can be effectively converted to improved BE throughput. I develop, implement, and evaluate a runtime system that achieves this goal with existing hardware. I leverage the cache partitioning, per-core frequency scaling, and thread masking of server processors. Evaluation results show the proposed solution enables 30% higher system throughput compared to solutions proposed in prior works while maintaining at least as good QoS for LC jobs. Second, I study resource contention in near-future heterogeneous memory architectures (HMA). This study is motivated by recent developments in non-volatile memory (NVM) technologies, which enable higher storage density at the cost of same performance. To understand the performance and QoS impact of HMAs, I design and implement a performance emulator in the Linux kernel that runs unmodified workloads with high accuracy, low overhead, and complete transparency. I further propose and evaluate multiple data and resource management QoS mechanisms, such as locality-aware page admission, occupancy management, and write buffer jailing. Third, I focus on accelerated machine learning (ML) systems. By profiling the performance of production workloads and accelerators, I show that accelerated ML tasks are highly sensitive to main memory interference due to fine-grained interaction between CPU and accelerator tasks. As a result, memory resource contention can significantly decreases the performance and efficiency gains of accelerators. I propose a runtime system that leverages existing hardware capabilities and show 17% higher system efficiency compared to previous approaches. This study further exposes opportunities for future processor architectures