These html pages are based on the PhD thesis "Cluster-Based Parallelization of Simulations on Dynamically Adaptive Grids and Dynamic Resource Management" by Martin Schreiber.
There is also more information and a PDF version available.

8.4 Scheduling decisions

The content and structure of this section is related to our work [SRNB13b] which is currently under review. The optimized target resource distribution D⃗ is computed based on the previously introduced data structures T⃗ as the specified optimization target and ⃗P as an per-application specified information.

We further drop the core dependencies of our original optimization function (8.2), yielding the simplified optimization function

⃗(i+1)           ⃗ (i) ⃗ ⃗
D     := foptimize(D   ,P,T ).

Optimizations are then applied with the constraints of all applications depending on P⃗ and the per-application optimization target T⃗.

Requirements on constraints: The constraints which are forwarded by resource-aware applications to the RM are then kept in ⃗P with one entry for each application. Then, the RM schedules the available resources based on the optimization target and these constraints. Here, we distinguish between local (optimizing resources for a single application) and global constraints (optimizing resources for multiple applications).

Local constraints: With constraints given by the range of cores between 1 and the maximum number of cores, an application can request a particular range of cores. These constraints make is challenging to optimize concurrently running applications since no knowledge on their performance state for a changing number of resources can be inferred and we refer to such constraints as local ones.

Global constraints: Such constraints can be evaluated by the optimization function in a way which optimizes the resources targeting at a global optimum of all applications.