D3.2. Mapping strategies

Release date: March 2015. Public.


Deliverable 3.2, “Mapping strategies”, covers the work done during the second phase of the project within WP3. The deliverable spans 12 months of work and handles the work done in task 3.2 "Mapping graph of parallel task dependencies to threads" to reach milestone MS2.

We describe the OpenMP characteristics that impacts on task-to-thread mapping: (1) The dependencies expressed by the extended static task depenendency graph (esTDG) derived by the P-SOCRATES compiler tool-chain (see deliverable D2.2.1) and (2) the implications that tied and untied clauses have on the schedulability of tasks.

We provide a lightweight run-time implementation that considers the esTDG. We evaluate the new run-time based on the memory consumption and the performance speed-up.

We identify the possible mapping strategies and evaluate the implications on the timing predictability of the system: (1) fully static mapping; (2) a dynamic mapping considering the current OpenMP run-time implementations (breadth-first scheduler of BFS and work-first scheduler or WFS).

Finally, we expose the implications that the timing analysis techniques (developed in WP3 and WP4) have to derive a time predictable mapping.

Access Document