Latency-aware Resource Management in Data Centres

Energy efficiency is a key issue in data centres. Data centres consume half of its maximum power even at low utilisation. In order to improve energy proportionality, machine utilisation is increased by co-locating best-effort (BE) workloads with latency-critical (LC) workloads. However, latency-critical workloads have strict quality-of-service (QoS) targets which must be met. When workloads are co-located, they share resources such as cores and last-level cache (LLC). A cluster manager is responsible for dynamically managing resources of the workloads in order to protect the performance of the LC workload while improving machine utilisation. This thesis aims to study an existing cluster manager called Intel PRM. Intel PRM uses cycles per instruction (CPI) a throughput based metric, to make resource management decisions when workloads are co-located. We aim to optimise the existing cluster manager by modifying it to make decisions based on the application-level latency. This thesis only deals with CPU resource management. We succeed in improving the throughput of the best-effort workload from 4.6% to 54.0% while providing 100% QoS-guarantee