Adaptive Power Shifting in Server Systems Ming Chen Xue Li
Introduction Power is one of the most important resources in data centers. CPU and main memory are two of the largest power gluttons. 85.4% 92.2% CPU & memCPU : mem
Motivation Power is a limited and precious resource for a server. CPU is NEITHER the only, NOR the largest consumer. Power needs to be controlled within a budget. Workload is varying in data centers. Over-provisioning may cause unnecessary cost. Power needs to be optimally allocated among server components, i.e., CPU and main memory. Goal: 1. Power control in the server-level. 2. Optimized allocation between CPU and main memory.
State of The Art Power control for CPU [Wang] in ISCA’09, main memory [Diniz] in ISCA’07 − Single component; Power control for servers [Lefurgy] in ICAC’07 − Assuming CPU is the major consumer Power shifting in server-level [Felter] in ICS’05 − Power is not controlled at the budget. − Performance is not optimal. Power control for racks [Wang] in HPCA’08, data centers [Wang] in PACT’09 − Complimentary to that work − Compose a complete power management scheme for data centers.
Who is Waiting? CPU mem CPU mem CPU mem CPU mem CPU mem CPU mem CPU is waiting.Memory is waiting.Coordinated. CPU: DVFS Memory: active memory size -> bandwidth
CPU-intensive workload Power should be shifted to CPU as much as possible. f opt = f max Memory-intensive workload Power should be shifted to memory as much as possible. f opt = f min Memory-intensive workload. f opt = ? What is the optimal allocation? For any piece of codes, it belongs to one of the following cases:
Proposed Solution − Ratio of the two weights: Memory Access Rate (MAR) − Optimization Weights the # of dispatched instructions in each control period the # of memory access requests in each control period Intuition − The more the # of activities is, the more power is required. Problems − MAR can guarantee the optimal allocation? − The weight ratio should be a function of MAR? Solution: − Power is controlled at the budget.
Schedules Idea verification (by the end of Sept) − Simulators integration; −Preliminary experiments to show the idea works; Controller design and tuning (by Oct 20) − Refinement of the idea − Refinement of sleeping scheme − Experiment plan Experiments and paper writing (until middle of Nov)