Thanunathan Rangarajan

Wireless Data Center Management: Sensor Network Applications and Challenges

The modern data centers (DCs) are essential to fulfilling ever-evolving computational demands around cloud computing, big data, and IT infrastructure. These DCs are facilities (Figure 1) that house computer systems and associated components such as networking and storage systems. To operate a DC, power supplies, network connections, environmental controls (e.g., air conditioning, humidity), and security infrastructure are needed. Technology and business challenges such as virtualization, load consolidation, real-time troubleshooting, and service-level guarantees require a robust and adaptive server management plan for enterprise. The majority of DC issues are related to overutilization of resources, application failures, data security, power usage effectiveness (PUE), and infrastructure costs. This requires proactive solutions that are business intelligent and built over a network of sense points that are guaranteed to deliver reliable trends and measurements in a reliable and timely fashion. Since it is expensive to build new DCs, the best option is to improve usage of an existing facility through lower infrastructure overhead to deliver better resource management. An optimal sensor network would perform real-time sensor-data collection and deliver a) improved server rack utilization, b) improved DC cooling, and c) improved loadbalancing through dynamic capping of thermally constrained systems.

Phase-aware predictive thermal modeling for proactive load-balancing of compute clusters

The increasing trend of high density computing environments have exacerbated the cooling infrastructure of the modern datacenters which contributes to mounting energy costs due to uncoordinated operation. By integrating information technology and infrastructure management through continuous monitoring, a balance between energy requirements of compute and cooling equipment can be achieved. Building an online thermal profile calculation with certain measure of accuracy is a complex problem due to the number of variables involved. In this paper we propose a phase-aware workload placement scheme that helps in reducing thermal variance in a cluster of compute nodes. We use a phase-aware machine learning approach to forecast server thermal profile which is then used for predicting the cluster-level thermal variance. We leverage Intel Xeon class server platform sensors and machine monitoring capability for fine grained assessment of power, thermal and compute utilization. We are able achieve thermal balance by applying intelligent placement algorithms by predetermining the thermal impact of a variation in workloads utilization on a prospective cluster of server using the forecasted temperature. Results from a prototype implementation on a typical server-cluster environment have demonstrated accurate thermal prediction and significant reduction in thermal variance.

A novel approach to determining compute headroom on a server platform

The ability to perform performance and power aware workload placement in the datacenter is vital to optimizing overall operational efficiency of the IT as well as the infrastructure itself. Traditional methods of measuring workload performance such as Operating System (OS) based performance counters do not deliver a consistent picture of the intricate relationship between compute resource consumption and its impact on the power and thermal profile of the compute node. As a result, the traditional methods constitute a barrier to effectively optimizing compute resource consumption in a holistic manner in the datacenter, perpetuating the management silos that exist between Information Systems and the supporting infrastructure, and impacting optimization of the primary cost drivers of a datacenter and the integration of performance sensitive applications. In this paper we present a new metric to correlate compute resource consumption with power to provide insights into the computational efficiency of the node for improved workload placement. We provide the theoretical background, followed by demonstrative examples and data to showcase the efficacy of this new approach. We finally conclude with specific recommendations based on this approach to help data centers achieve resource optimization.

Systemic optimization of on-chip thermoelectric cooling

Economic viability of on-package, in-situ cooling based on Thin-film Thermoelectric Coolants (TF-TEC) for hot-spot cooling involves myriad challenges necessitating engineering trade-offs. Principal factors include the cost of integration, the net energy consumption of the TEC based system, as well as system-level complexities arising from issues such as mutual thermal conflicts and interdependencies between the TEC and other package-level entities such as the Thermal grease (TIM), impact on the external convective cooling system, and the number of hot-spots present. In this paper, we examine these challenges both analytically and empirically, and propose a heuristic based method to overcome them. The method forms the basis for a generic optimization framework that enables system-level optimization of on-chip thermoelectric cooling in a commercial microprocessor package. We find a resultant cooling of up to 3°C at TDP delivered per core with a ~11% improvement in energy efficiency.