01.
06.
2015
Simulation of a Scheduling Algorithm for DAG-based Task Models
Maksym Planeta
TU Dresden
Sondertermin: 13:30 Uhr, APB 3105
Efficiency of a parallel applications merely depends on a scheduling algorithm which distributes sequential portions of application to multiprocessor systems. State of the art algorithms can be divided in following two classes. Just-in-time (JIT) algorithms are simple algorithms, which can schedule almost any application and have minimal requirements. Full-ahead algorithms are more computation intensive and require more information about a parallel program, but they are able to build more effective schedules.
Directed acyclic graph (DAG) is a typical instrument to model structure of a parallel application, which is used by full-ahead algorithms. DAGs show dependencies between sequential parts of an application and communication between them.
This thesis presents a Horizon algorithm class for scheduling DAG-based parallel applications. The Horizon algorithm extends a JIT algorithm class by providing some restricted knowledge about parallel program structure. With this new knowledge the Horizon algorithm manages to outperform a JIT algorithm. To the best of my knowledge this is the first attempt to fill the niche between full-ahead and JIT algorithms.
|
