Source: PowerNap: Eliminating Server Idle Power (PDF)
By 2011, U.S. data centers will consume 100 billion kWh at a cost of $7.4 billion per year (U.S. EPA, “Report to congress on server and data center energy efficiency,” U.S. Environmental Protection Agency, Tech. Rep., Aug. 2007.). Oofda.
At idle, current servers still draw about 60% of peak power. (L. Barroso and U. H ¨
olzle, “The case for energy-proportional computing,” IEEE Computer, Jan 2007.) (X. Fan, W.-D. Weber, and L. A. Barroso, “Power provisioning for a warehouse-sized computer,” in Proc. of the 34th Annual International Symposium on Computer Architecture, 2007.) and (C. Lefurgy, X. Wang, and M. Ware, “Server-level power control,” in Proc. of the IEEE International Conference on Autonomic Computing, Jan 2007.)
In other words this is an area of inefficiency that's ripe for improvement and with the right technology you can capture some value out of improving efficiency. Energy that isn't required to be consumed, does not have to be produced. Avory Lovins popularized this idea with the term "Negawatts" which are the Megawatts that do not have to be produced. Energy that isn't produced does not cause natural gas or coal to be burned.
Ideally they want to simply turn off idle systems. However their analysis shows that with typical server loads idle periods are less than a second. With current systems it's impractical to for a system to sleep in such short intervals.
Data center operators must provision their systems for peak loads rather than average loads. That means there will always be excess server capacity.
An existing technique is dynamic voltage and frequency scaling (DVFS) which almost entirely eliminates CPU power usage during idle periods. However the CPU has a small part of total server energy consumption (other components like memory or disks etc also consume power). Improving DVFS algorithms is an area of active research.
One idea is 'Energy proportional computing' which would extend DVFS to the entire system. However not every component can be redesigned to vary their power consumption to match utilization.
PowerNap is simpler because it requires components to only implement two modes: active, and napping (with minimized power use)
Transition speed between napping and active modes is the dominant factor in determining the power savings potential and response time impact of PowerNap. The response time is a key factor in server systems as customers of the service run on the server demand appropriate response time. Think of a web server and the concept of "slow" which really means it took the page too long to render. The customer of a web site doesn't care if a napping server causes slow response time, they want to see the web page now.
Hardware mechanisms they depend on are:
- ACPI S3 "sleep" state
- DRAM Self-refresh which can make cut power requirements during napping periods to 4% of power requirements during active periods.
- Solid State Disks.. consume miniscule amounts of power when idle and do not need to transition to a sleep state.
- Wake-on-LAN causes the system to turn on when a packet arrives on the network port.
- Environmental monitoring and service processors
- Variable speed fans that can turn themselves down (or off) when required
- Power provisioning (RAILS) explained further on
Current server power systems lead to wasted energy. A new system, RAILS, is proposed to decrease that waste. The idea is to use multiple small power supplies, the small power supplies each are built to operate efficiently at low power levels, and the small power supplies are dynamically turned on and off as required. The individual power supplies are each attached to a common power bus, and each server is also attached to the same power bus. The power supplies have to be sized so each power supply gives enough power for one server blade's power requirements.
]
No comments:
Post a Comment