Run-Time Resource Management at the Operating System Level - Multi-objective Design Space Exploration of Multiprocessor SoC Architectures - page 138

6.5.2 Integration Requirements and Goals

The idea to integrate effectively the DSE techniques within an Operating System
cannot disregard the problems highlighted in the previous section. Moreover, to
provide a general control solution which is acceptable by the Linux community, it is
required to integrate the DSE control with the existing frameworks.
To satisfy all these goals, by providing a flexible and efficient OS integration of
the DSE control techniques, a hierarchical design should be considered. In such a
design, the monitoring, management, control and optimization strategy is operated at
different granularity levels. Each granularity level collects requirements from higher
level, runs a specific optimization policy, and finally identifies a set of constraints
delivered to lower levels.
Such a hierarchical approach requires the development of a new framework, in be-
tween the DSE control policy and the in-kernel existing frameworks, which allows to
collect application requirements and match them against resources availability. This
matching is performed based on a dynamic optimization policy to be developed at
different abstraction levels and according to different optimization goals. Run-time
optimization policies derived from DSE are considered as coarse grained configu-
ration points, related to the specific off-line profiled application. Nevertheless, such
information could be aggregated at run-time with other system requirements in order
to achieve a fine-grained and system-wide optimization.
The main abstraction levels are user-space, kernel-space and device-space. The
user-space level corresponds to the DSE control policy, which can be defined off-
line for the fine tuning of profiled critical applications. An example of those kind of
policies is presented in [ 14 ]. The kernel-level control policy is defined by the new
framework, which allows to collect and aggregate requirements from both critical and
best-effort applications. This level is somehow similar to the cross-layer frameworks
presented in Sect. 6.4.2 . Finally, the device-space level is related to each pre-existing
subsystem specific control policy like those presented in Sect. 6.4.1 .
Regarding the optimization goals , the definition of a new kernel-space frame-
work, which could have a system-wide view on run-time system state and resource
availability, allows more easily to develop a control policy which is aware of run-
time phenomena. This new framework could provide support for optimization goals
which are difficult to define off-line, such as resource usage fairness, application
performance, power consumption and thermal management.
The new RTRM kernel-space framework proposed for the integration of the DSE
defined control policies allows to effectively target two main goals: dynamic resource
partitioning and resource abstraction.

6.5.2.1 Dynamic Resource Partitioning

Each application could be associated to a certain “priority level” related to the differ-
ent impact that the application (either critical or best-effort) could have on the overall
user experience. Thus, the RTRM framework should provide a support to handle both: