Specifications

Chapter 2. Architecture and technical overview 71
Draft Document for Review May 12, 2014 12:46 pm 5102ch02.fm
Power saver mode is not supported during system startup, although it is a persistent
condition that is sustained after the boot when the system starts executing instructions.
򐂰 Dynamic power saver mode
Dynamic power saver mode varies processor frequency and voltage based on the
utilization of the POWER8 processors. Processor frequency and utilization are inversely
proportional for most workloads, implying that as the frequency of a processor increases,
its utilization decreases, given a constant workload. Dynamic power saver mode takes
advantage of this relationship to detect opportunities to save power, based on measured
real-time system utilization.
When a system is idle, the system firmware lowers the frequency and voltage to power
energy saver mode values. When fully utilized, the maximum frequency varies, depending
on whether the user favors power savings or system performance. If an administrator
prefers energy savings and a system is fully utilized, the system is designed to reduce the
maximum frequency to about 95% of nominal values. If performance is favored over
energy consumption, the maximum frequency can be increased to up to 111.3% of
nominal frequency for extra performance.
Dynamic power saver mode is mutually exclusive with power saver mode. Only one of
these modes can be enabled at a given time.
򐂰 Power capping
Power capping enforces a user-specified limit on power usage. Power capping is not a
power-saving mechanism. It enforces power caps by throttling the processors in the
system, degrading performance significantly. The idea of a power cap is to set a limit that
must never be reached but that frees extra power that was never used in the data center.
The
margined power is this amount of extra power that is allocated to a server during its
installation in a data center. It is based on the server environmental specifications that
usually are never reached because server specifications are always based on maximum
configurations and worst-case scenarios.
򐂰 Soft power capping
There are two power ranges into which the power cap can be set: power capping, as
described previously, and soft power capping. Soft power capping extends the allowed
energy capping range further, beyond a region that can be guaranteed in all configurations
and conditions. If the energy management goal is to meet a particular consumption limit,
then soft power capping is the mechanism to use.
򐂰 Processor core nap mode
IBM POWER8 processor uses a low-power mode called
nap that stops processor
execution when there is no work to do on that processor core. The latency of exiting nap
mode is small, typically not generating any impact on applications running. Therefore, the
IBM POWER Hypervisor™ can use nap mode as a general-purpose idle state. When the
operating system detects that a processor thread is idle, it yields control of a hardware
thread to the POWER Hypervisor. The POWER Hypervisor immediately puts the thread
into nap mode. Nap mode allows the hardware to turn the clock off on most of the circuits
in the processor core. Reducing active energy consumption by turning off the clocks
allows the temperature to fall, which further reduces leakage (static) power of the circuits
causing a cumulative effect. Nap mode saves 10 - 15% of power consumption in the
processor core.
򐂰 Processor core sleep mode
To be able to save even more energy, the POWER8 processor has an even lower power
mode referred to as
sleep. Before a core and its associated private L2 cache enter sleep
mode, the cache is flushed, transition lookaside buffers (TLB) are invalidated, and the
hardware clock is turned off in the core and in the cache. Voltage is reduced to minimize