Data Center, Power Management, Dynamic Voltage and Frequency Scaling, Energy Efficiency, 데이터센터, 전력 관리, 동적 전압 및 주파수 조절, 에너지 효율성
Table Of Contents
I. Introduction 1 1.1 Contributions 2 1.2 Organization 4 II. Background 5 2.1 Network Packet Processing and New API (NAPI) 5 2.2 Processor Power Management 6 2.3 Hypervisor Power Management 8 2.4 Hypervisor Scheduling and Performance Isolation 9 2.5 P-states by CPU utilization on Hypervisor 10 III. Related Work 12 3.1 Power Management under SLO constraints 12 3.2 Power Management with Virtualization 13 3.3 Power Management exploiting Uncore Frequency Scaling 14 IV. NMAP: Network packet processing Mode Aware Power management 15 4.1 Power Management for Latency Critical Workloads 15 4.2 Motivation 17 4.2.1 Network Packet Processing Mode Transition of NAPI and Network Load 17 4.2.2 Limitation of CPU utilization based Power Management 18 4.3 NMAP: Network Packet Processing Mode Aware Power Management 19 4.3.1 NMAP Architecture 21 4.3.2 NMAP Exploiting Network Packet Processing Mode Transition 22 4.4 Discussion 25 4.4.1 Limitation of Power Management requiring short latency of V/F Transition 25 4.4.2 Impact of Sleep State on Latency-Critical Workloads 28 4.5 Evaluation 30 4.5.1 Methodology 30 4.5.2 Comparison with Conventional Power Management 31 4.5.3 Comparison with state-of-the-arts, SLO-aware Power Managements 34 4.6 Summary 37 V. VIP: VIrtual Performance-State for Power Management of Virtual Machines 38 5.1 Hypervisor based Virtualization with Power Management 38 5.2 Experimental Methodology 39 5.3 Limitation of current per-core based Power Management 41 5.3.1 Static Power Management 41 5.3.2 Dynamic Power Management with Heterogeneous Virtual Machines 42 5.4 VIP Architecture 44 5.4.1 Supporting Per-VM Power Management 45 5.4.2 Hypervisor-assisted Per-VM Power Management 46 5.4.3 Hypervisor Scheduling with VIP 50 5.4.4 Discussion 51 5.5 Evaluation 52 5.5.1 P-state Distribution 53 5.5.2 Performance and Energy Consumption 54 5.5.3 Performance Comparison of Hypervisor Scheduling 55 5.6 Summary 56 VI. CoScale: Core/Uncore Frequency Scaling for Consolidated Workloads 58 6.1 Power Oversubscription with Power Capping 58 6.2 Motivation 59 6.2.1 Characteristics according to Core and Uncore Frequency 59 6.2.2 Performance Impact of Uncore Frequency Scaling 61 6.2.3 Challenge of Core/Uncore Frequency Scaling for Consolidated Workloads 62 6.3 Architecture 64 6.3.1 Core/Uncore Frequency Scaling by Bayesian Optimization 64 6.3.2 CoScale: Design and Implementation 66 6.4 Evaluation 69 6.4.1 Methodology 69 6.4.2 Experimental Results 71 6.5 Summary 73 VII. Conclusion 75 References 76