Publications

This work studies whether AI data centers can behave as flexible grid resources for demand response programs. It focuses on the gap between theoretical load flexibility and what is actually available once AI workload deadlines, performance targets, infrastructure constraints, and operational risk are considered. The paper investigates how much power consumption can be shifted or reduced, what kinds of workloads are better suited to participation, and how data center operators might expose flexibility to the grid while protecting service quality. The broader goal is to connect AI infrastructure operations with power-system needs in a way that is measurable and actionable.

This paper reports lessons from anomaly detection in Chameleon Cloud, emphasizing the realities of deploying detection methods in an operational research cloud rather than a controlled benchmark. It discusses the data quality, labeling, workload diversity, and infrastructure variability issues that shape anomaly detection performance in practice. The work is useful for understanding why cloud reliability tools need to account for changing usage patterns, noisy telemetry, and the difference between statistically unusual behavior and operationally meaningful incidents.

This paper asks whether data center demand response is a real opportunity for sustainable computing or mostly a conceptual promise. It examines how data centers could adjust power consumption in response to grid needs, while accounting for workload constraints, service-level expectations, power delivery limits, and the economics of participation. The work positions demand response as a systems problem: useful grid flexibility depends not only on available electrical capacity, but also on scheduling policies, workload tolerance, market rules, and operator willingness to trade compute immediacy for energy-system value.

vCap presents adaptive power capping for virtualized servers, where multiple virtual machines or applications share a physical platform under a power budget. The paper addresses the challenge of enforcing caps while minimizing performance degradation for workloads with different resource needs and scaling behavior. It proposes runtime mechanisms that adjust power and resource decisions dynamically, allowing the system to respond to workload changes instead of applying a fixed cap uniformly. The work is part of a broader push toward energy-aware cloud and server management.

This paper studies energy-efficient server consolidation for multi-threaded cloud applications. It looks beyond simple utilization-based consolidation by considering how multi-threaded applications scale and interfere when placed together. The goal is to reduce the energy cost of running cloud workloads while preserving acceptable performance, which requires deciding when consolidation saves power and when it creates contention that offsets the benefit. The work contributes to runtime and placement strategies for making cloud infrastructure more energy proportional.

This paper connects server-level power capping with the possibility of data center participation in power markets. It studies how dynamic power control can make data centers more responsive to electricity prices or market signals, while still serving computational workloads. The work treats data centers as active energy-market participants rather than passive electricity consumers, requiring coordination between workload management, server power limits, and market-driven operating decisions. This theme anticipates later work on grid-interactive and demand-responsive computing infrastructure.

This paper investigates how co-scheduling parallel workloads can reduce the energy cost of computing. The central observation is that the energy efficiency of a workload depends on what else is running, how resources are shared, and whether the system is operating at an efficient point. By coordinating which parallel jobs run together, the system can improve utilization and reduce wasted energy while maintaining throughput. The work contributes scheduling techniques for energy-aware high-performance and server-class computing environments.

This paper presents initial case studies on software optimization across performance, energy, and thermal distribution. Rather than treating energy or temperature as after-the-fact hardware concerns, it explores how software choices influence system-level behavior. The case studies help show that performance optimization alone may not produce the best energy or thermal outcomes, and that software can be tuned with multiple objectives in mind. The work is an early contribution to holistic optimization for green computing systems.

This paper studies how to identify energy-efficient operating points for parallel workloads. It considers the tradeoff between running faster with more resources and running more efficiently at a lower-power point, recognizing that the best choice depends on workload scaling and system behavior. The work provides methods for locating operating points that reduce energy while maintaining useful performance, contributing to runtime and design-time strategies for energy-aware multicore and parallel computing systems.

Pack & Cap combines adaptive dynamic voltage and frequency scaling with thread packing for workloads operating under power caps. The paper addresses the problem that simply lowering frequency or uniformly limiting power can leave performance on the table. By coordinating where threads run and how cores are scaled, the system can better satisfy a power cap while preserving throughput. The work is significant for power-constrained multicore systems because it treats placement and frequency control as coupled decisions rather than independent knobs.