Adaptive Power Management

From RAD Lab

Vision

• Vision Paper

Improving efficiency of hardware components

• Google: towards more efficient computing

Price of electricity

• all larger data centers in northern California are charged for electricity using the E-19 or E-20 rate schedules; more details here
• demand response programs, to get more details, go to the tariff book (link below) and search for the following tariff names: E-BIP, E-OBMC, E-SLRP, E-DBP, E-POBMC, E-CPP
• tariff book

Data center power consumption

• Benchmarking Data Centers (work done at LBL): [1]
  • energy use breakdown: [2]
  • case study reports
  • Overview of LBL datacenter power measurement activities
• EnergyStar initiatives for datacenter power management

Related work

Server-level stuff

• The case for power management in web servers by Bohrer, Elnozahy et al.
• Managing adaptive server resources... by Chase et al.
• Pinheiro at al, Rutgers - turn off whole servers when possible. Most related to our ideas.

Thermal load-balancing

• main idea: by either a) moving workload between servers, or b) directing cold air from AC units to hot parts of a data center, reduce the maximum temperature in the room. If we decrease the max temperature by (say) 3 degrees Celsius, we can increase the temperature coming from the AC units (by 3 degrees Celsius) and still keep the room cool. This way the AC units operate in a more efficient regime and save power.

Dynamic Voltage Scaling in servers

• since most servers in data centers are underutilized, we could reduce the voltage/frequency of CPUs to reduce power consumption

References:

• Ensemble-level Power Management for Dense Blade Servers (ISCA '06): [3]

Heterogeneous Chip Multiprocessors

Heterogeneous Chip Multiprocessors contain multiple cores, each having different capabilities and performance levels: the "smaller" cores are slower, but more power efficient, while the "bigger" cores are faster, but consumer more power. The authors propose to dynamically switch between cores depending on the application requirements; their evaluation shows that this technique could achieve significant power savings while sacrificing only little in performance.

References:

• [4] (ISCA '04)
• [5] (IEEE Computer, November '05)