Environmental monitoring using wireless sensors plays a key role in detecting hotspots or over-cooling conditions in a data center (DC). Despite a myriad of Data Center Wireless Sensor Network (DCWSN) solutions in literature, their adoption in DCs is scarce due to four challenges: low reliability, short battery lifetime, lack of adaptability, and labour intensive deployment. The main objective of this research is to address these challenges in our specifically designed hierarchical DCWSN, called Low Energy Monitoring Network (LEMoNet). LEMoNet is a two-tier protocol, which features Bluetooth Low Energy (BLE) for sensors communication in the first tier. It leverages multi-gateway packet reception in its second tier to mitigate the unreliability of BLE. The protocol has been experimentally validated in a small DC and evaluated by simulations in a midsize DC. However, since the main application of DCWSNs is in colocation and large DCs, an affordable and fast approach is still required to assess LEMoNet in large scale. As the first contribution, we develop an analytical model to characterize its scalability and energy efficiency in a given network topology. The accuracy of the model is validated through extensive event-driven simulations. Evaluation results show that LEMoNet can achieve high reliability in a network of 4800 nodes at a duty cycle of 15s. To achieve the network adaptability, we introduce and design SoftBLE, a Software-Defined Networking (SDN) based framework that provides controllability to the network. It takes advantages of advanced control knobs recently available in BLE protocol stacks. SoftBLE is complemented by two orchestration algorithms to optimize gateway and sensor parameters based on run-time measurements. Evaluation results from both an experimental testbed and a large-scale simulation study show that using SoftBLE, sensors consume 70% less power in data collection compared to those in baseline approaches while achieving the Packet Reception Rate (PRR) no less than 99.9%. One of its main steps of DCWSN commissioning is sensor localization, which is labour-intensive if is driven manually. To streamline the process, we devise a novel approach for automated sensor mapping. Since Radio Frequency (RF) alone is not a reliable data source for sensor localization in harsh and multi-path rich environments such as a DCs, we investigate using non-RF alternatives. Thermal Piloting is a classification model to correlate temperature sensor measurements with the expected thermal values at their locations. It achieves an average localization error of 0.64 meters in a modular DC testbed. The idea is further improved by a multimodal approach that incorporates pairwise Received Signal Strength (RSS) measurements of RF signals. The problem is formulated as Weighted Graph Matching (WGM) between an analytical graph and an experimental graph. A parallel algorithm is proposed to find heuristic solutions to this NP-hard problem, which is 30% more accurate than the baselines. The evaluation in a modular DC testbed shows that the localization errors using multi-modality are less than one-third of that of using thermal data alone. / Thesis / Candidate in Philosophy

With the rapid growth of servers and applications spurred by the Internet economy, power consumption in today's data centers is reaching unsustainable limits. This has led to an imminent financial, technical and environmental crisis that is impacting the society at large. Hence, it has become critically important that power consumption be efficiently managed in these computing power-houses of today. In this work, we revisit the issue of adaptive power and performance management of data center server platforms. Traditional data center servers are statically configured and always over-provisioned to be able to handle peak load. We transform these statically configured data center servers to clairvoyant entities that can sense changes in the workload and dynamically scale in capacity to adapt to the requirements of the workload. The over-provisioned server capacity is transitioned to low-power states and they remain in those states for as long as the performance remains within given acceptable thresholds. The platform power expenditure is minimized subject to performance constraints. This is formulated as a performance-per-watt optimization problem and solved using analytical power and performance models. Coarse-grained optimizations at the platform-level are refined by local optimizations at the devices-level namely - the processor & memory subsystems. Our adaptive interleaving technique for memory power management yielded about 48.8% (26.7 kJ) energy savings compared to traditional techniques measured at 4.5%. Our adaptive platform power and performance management technique demonstrated 56.25% energy savings for memory-intensive workload, 63.75% savings for processor-intensive workload and 47.5% savings for a mixed workload while maintaining platform performance within given acceptable thresholds.