Run-Time Resource Management at the Operating System Level - Multi-objective Design Space Exploration of Multiprocessor SoC Architectures - page 139

critical workloads, that have hard requirements in terms of resource usage along with
execution behaviors, and best-effort workloads, for which a penalty either does not
impact on perceived behaviors or produce a tolerated QoS degradation.
Both classes of applications should access the available resources through a single
run-time resource manager framework. Thus, the role of this framework is to account
for available resources, and grant access to these resources to demanding application
according to their priority level. To efficiently manage this scenario, the resources
could be dynamic partitioned, taking into consideration current QoS requirements
and resources availability. This dynamic partitioning will allow to grant resources to
critical workloads while dynamically yield these resources to best-effort workloads
when (and only while) they are not required by critical ones, thus optimize resource
usage and fairness.

6.5.2.2 Resource Abstraction

To suitably tackle the run-time phenomena problem, the RTRM kernel-space frame-
work should handle a decoupled perspective of the resources between the users and
the underlying hardware. The user applications will see virtual resources but they
will not be aware of which of the physical resources are effectively available. At
run-time the RTRM will perform the virtual-to-physical mapping according to the
current objective function (low power, high performance, etc.) and run-time phe-
nomena (process variation, temporal and spatial temperature gradients, hardware
failures and, above all, workload variation).

6.6 Conclusions

The aim of this chapter is to provide the overall picture of the components and
goals of a system-wide resource manager. The need for a Run-Time Resource Man-
ager(RTRM) is dictated by the non-deterministic nature of the complex modern
applications. This need, in addition, poses new challenges in finding out a low-
overhead solution that satisfies all the functional and non-functional requirements.
From a performance and QoS point of view, having multiple applications means
having at any point of time different perspectives of the entire system.
Our current research takes full advantages from the DSE-based design flow devel-
oped within the MULTICUBE project. The coarse grain analysis of the best operating
points to be used at run-time will be improved and integrated in a more comprehen-
sive framework, working at the level of the Operating System, capable to provide a
fine-grain tuning of the system configuration considering a system-wide optimization
perspective. Our research goal is to develop a hybrid centralized/distributed approach,
conveying into a hierarchical strategy for managing power consumption/energy re-
quirements. A hierarchical strategy has the benefits of gathering local control and
centralized optimal solution to the problem. In addition, such a hierarchical control