Hardware Reference
In-Depth Information
problem. A working mode of a device can also be represented in the space of these
metrics by identifying a corresponding device working region (DWR).
The model layer exploit these abstraction information to automatically build a
representation of all the system-wide feasible configurations (FSCs), each one iden-
tifying a working points of the entire system where a certain QoS level can be
granted.
This model is suitable for supporting an efficient global optimization strategy
provided by the optimization layer . This policy could support a multi-objective
optimization strategy defined on the considered metrics.
The framework provides the implementation of a global optimization policy which
rely on Linear Programming to identify a solution-equivalent and efficient optimiza-
tion strategy. This strategy is based on three main tasks: FSC Identification, FSC
Ordering and FSC Selection.
The FSC Identification exploits the information provided by the abstraction layer.
This layer provides a multi-dimensional solution space, defined by the optimization
metrics (PSM/ASM), and the device working regions (DWR) defined by each device.
These information are used in the model layer to automatically identify the set of
Feasible System-Wide Configurations (FSC) that identifies a platform independent
representation of the system resources and capabilities.
In the FSC Ordering task, the global optimization policy exploits the FSC-based
abstract system view provided by the model layer, to order the FSC previously
identified according to the multi-objective optimization goal. Each time the system
use-case change, the optimization goal is updated, and thus the FSC should be
re-ordered.
Application requirements are collected and translated into constraints which inval-
idate unfeasible FSCs. The optimal FSC is selected, considering both the previously
identified ordering and their validity defined by the current constraints, by running
the FSC Selection task.
Each task has different run-time overhead and activation frequency. The FSC
identification is the more complex task, it is required just at system-boot. Instead,
FSC selection must run each time an application assert a requirement but it has a
negligible impact thanks to the support provided by the previous tasks.
6.5
Exploiting DSE to Support RTRM
The Design Space Exploration (DSE) techniques have been demonstrated to be a
valuable tool for the exploration and optimization of hardware platforms. However,
their usage at run-time for the optimization of software behaviors has been only
recently explored [ 14 ]. Even if the approach is promising, the integration of a DSE
based optimization policy for the power optimization of a computing platform could
exhibit less flexibilities. This is especially true if we consider the new generation
systems, based on a deep sub-micron production processes, which run multiple
applications on top of an Operating System.
Search WWH ::




Custom Search