RAAM consults a variety of situational information
about the state of the system, including performance
measurements and models of system performance as a
function of resources. The simplest of this information
involves updates from performance counters to reflect
message volume across channels or energy usage. Another source of information is periodic performance reports from the application [12, 14]. A performance report contains Cell-specific performance metrics (e.g., the
amount of progress made on application-specific deadlines). These reports can be combined with applicationsupplied information about what a meaningful unit of
work would be and how often it must be completed (i.e.,
frames/second) to adjust allocations.
We are investigating a variety of techniques for resource adaptation. Although one could blindly increase
the resources given to a Cell until performance goals are
met, such a strategy is unlikely to work for complex con
figurations of Cells nor is it capable of incorporating subtleties in the level of importance of some Cells over others. Instead, we believe that two components are necessary: first, accurate models of the performance of Cells
(or the change in performance) as a function of resources
and second, a framework in which to drive the juggling
of resources among Cells. One framework that we are
investigating is a form of convex optimization over resources that attempts to minimize an “urgency” derived
from the degree to which Cells miss their deadlines