ECM 1413 Energy Efficient Computing Flashcards
The energy problem:
our sources of energy are steadily decreasing, whilst our demand for energy is steadily increasing
the emission of greenhouse gases is now seen
by most climatologists as being linked to global warming
Rising Electricity demand by 2030
Electricity demand for information and communications technologies (ICT) might rise to 8% of total electricity demand by 2030, in the best case, or 21% in the worst case
The technology developed by Intel, Microsoft, to combat high energy usage on current-day systems
The Advanced Configuration and Power Interface (ACPI)
ACPI States
1 per-system power states G0-G3
2 per-device (disks, keyboards, etc.) power states D0-D3
3 per-processor power states C0-C3 (substates of G0)
4 per-device/processor performance states P0-P15 (substates of D0/C0)
The G0, D0, C0, and P0 states have normal power consumption; others have successively lower power consumption.
Global State G0 Name
Powered on
Global State G1 Name
Sleeping
Global State G2 Name
Powered off
Global State G3 Name
Unplugged
Sleep State S0 Description
Standard operation
Sleep state S1 Description
CPU stops running. CPU caches are shut down. Lower power states of hardware devices
Sleep state S2 Description
CPU Powered off
Sleep state S3 Description
Shut down more devices. Makes it harder for devices to wake the system
Sleep state S4 description
Save memory to disk and shut down memory
Device state D0 Description
Fully operational
Device states D1, D2 description
Intermediate states (optional)
Device state D3 Description
Device is off
Processor State C0 Description
Fully operational
Processor State C1 Description
The processor stops executing instructions. Caches are maintained. Short wake time. Must wake when receiving an interrupt
Processor states C2, C3 Description
Similar to C1 but deeper sleep and longer wait time (optional)
Performance States P0 -> P15
A processor could save energy by scaling down the power and clock frequency (P0 ⇒ P1, . . . , P15).
Device power states differ in terms of
Power consumption
Device context retainment
Device driver behavior
Restore time Wake-up capability
Most transitions go to or from D0.
Magnetic Disks saving energy
A magnetic disk could save energy by
turning off spindle motor (D0 ⇒ D1)
turning off controller (D0 ⇒ D3)
Keyboards saving energy
A keyboard could save energy by
only providing wake functionality, turning off indicator lamps, while preserving device context such as the state of lock keys (D0 ⇒ D1)
turning off when not needed (D0 ⇒ D3)
Processor states are applied to individual cores
Synergy effects occur when…
… shutting down multiple cores within the same domain/processor/computer since you then can also shut down resources that these cores share
Smartphones have both high-performance cores and energy-saving cores, and can apply different processor states to these
Traditional Cloud Data Centres PUE
Traditional cloud data centres using off-the-shelf components have a Power Usage Effectiveness (PUE) of 2.0, far from a perfect 1.0.
Hyperscale Cloud Data Centres
The Open Compute Project is driven by large companies that have an interest in building more efficient data centres.
The project shares designs and best practices on data centres at no cost.
Hyperscale cloud data centres using Open Compute Project components have a Power Usage Effectiveness (PUE) of 1.2, close to a perfect 1.0.
One Open Compute Project server replaces 3.75 traditional ones.
Among the ways to lower power usage effectiveness are:
1 bespoke hardware
larger processors, more memory, more disk storage, remove graphics cards, lights, etc.
2 smarter cooling
use locations with a cold climate
3 mobile computing
let mobile devices do a larger share of the work
