Towards Energy-Efficient and Real-Time Cloud Computing

Abstract

Cloud data centers have become the backbone of modern computing infrastructure, supporting an ever-expanding range of applications from enterprise workloads to real-time services. However, this growth has led to unprecedented energy consumption, with data centers now accounting for approximately 2% of global electricity usage. Energy costs rep- resent the largest operational expense for data center operators, while the environmental impact of this consumption poses significant sustainability challenges. This dissertation ad- dresses the critical need for energy-efficient cloud resource management through the design, implementation, and evaluation of two complementary frameworks: EGRET and VMaestro. EGRET (Energy-efficient Gradual Real-time Execution Tuning) introduces a novel ap- proach to IT-side energy optimization by seamlessly integrating dynamic voltage and fre- quency scaling (DVFS) with virtual machine (VM) consolidation for real-time cloud work- loads. Unlike traditional consolidation techniques that focus solely on packing efficiency, EGRET employs a frequency-aware placement strategy that minimizes the global increase in CPU frequencies caused by VM migrations while ensuring real-time deadlines are met. Experimental evaluation using realistic cloud workload traces demonstrates that EGRET achieves 41.6% IT energy reduction compared to static-frequency baselines while maintain- ing service-level agreement (SLA) compliance comparable to existing approaches. Building upon EGRET’s foundation, VMaestro extends energy optimization to the fa- cility level by explicitly modeling thermal dynamics and integrating cooling system control. VMaestro addresses a fundamental limitation of IT-only approaches: the failure to account for the substantial energy consumed by cooling infrastructure, which typically represents 30–40% of total facility power.