Run-Time Resource Management at the Operating System Level - Multi-objective Design Space Exploration of Multiprocessor SoC Architectures - page 131

6.4.2 Cross-Layer Techniques

Mechanisms and techniques supporting power management can be implemented at
different abstraction levels; not only at architectural level but also at software level.
Applications are aware of their Quality-of-Service (QoS) expectations. For instance,
if we consider the playback of a network video stream: then we could easily identify
at the application level some of the requirements, e.g., in terms of network bandwidth
and decoding processing workload. Thus, the development of holistic approaches
should support the aggregation of data from multiple layers into power management
decisions. Cross-Layer techniques try to exploit mechanisms from different abstrac-
tion levels at the same time. The idea behind them is to provide properly defined
mechanisms to collect abstract information from the higher abstraction levels, i.e.,
user-space applications, and exploit them to give some useful hints to the lower
abstraction level techniques in order to improve the exploitation of the available ar-
chitectural mechanisms. Cross-layer techniques aggregate data from multiple layers
into power management decisions; indeed a properly defined interface allows the
user space to assert Quality-of-Service requirements and exploit these information
to support system-wide optimization. These techniques could be further grouped into
two categories, centralized and distributed . Centralized techniques have been devel-
oped mainly to support the power optimization of relatively simple and dedicated
embedded systems (e.g. personal media player), but these have some scalability
problems related to their complexity which impact on the implementation effort. On
the other hand, distributed techniques are designed to be more scalable to easily
address much more complex architectures (e.g. new generation smart-phones).
The power optimization techniques proposed in this class are essentially based
on the definition of a single coordination entity, which stands in between the user-
space applications and the available architectural mechanisms. However, we could
identify essentially two orthogonal approaches: centralized and distributed; the main
difference is in the role of the coordination entity. In centralized approaches the
coordination entity has a direct control on the available mechanisms which are used
to perform power management according to a single and system-wide optimization
policy driven by the requirements collected from user-space. Distributed approaches,
instead, not only implement a lightweight, single, and system-wide optimization
policy but also exploit many other devices and subsystem specific policies. The
idea is to implement a distributed control model where user-space requirements are
aggregated and used to feed some input to more specialized local controls.

6.4.2.1 Dynamic Power Management

The Dynamic Power Management software (and hardware, refer to architecture,
presented in [ 3 ]) is both an architectural and interface proposal for a centralized
cross-layer technique targeting high-performance embedded systems. Purpose of this
proposal is to exploit effective power management mechanisms from the architectural
view-point and from the management view-point at the OS level. This framework