TY - JOUR
T1 - Multi-layer task graph clustering for mapping and scheduling onto embedded systems
AU - Arafeh, Bassel R.
PY - 2012/9
Y1 - 2012/9
N2 - This work is motivated by the general trend that embedded systems design has high tolerance to longer compilation time. Based on this added flexibility in compilation, this work proposes a multilevel (multi-layer) graph partitioning framework for solving the mapping and scheduling problem for distributed heterogeneous embedded systems. However, this work concentrates on developing a multi-layer task graph clustering scheme as a part of a multi-step scheduling approach. This work introduces linear clustering techniques for vertex matching in Directed Acyclic Graphs (DAGs) that enable the construction of multi-layer coarsened DAGs with different granularity levels. This paper addresses the major obstacle facing the DAG contraction process, which is the formation of cycles, and discusses this issue for direct and chained dependence cycles. Solutions are provided that establish the basis for performing cycle testing in order to avoid the formation of cycles under certain conditions. This research shows, by analysis, that the upper-bound time complexity of the proposed cycle testing method is O(|V|3), where V is the set of task nodes. Also, this research shows, by analysis, that the complexity of the proposed multi-layer DAG coarsening scheme is O(|V|6). The simulation results verify the cost-effectiveness of the proposed scheme, and show its capabilities in generating clusters at different granularities.
AB - This work is motivated by the general trend that embedded systems design has high tolerance to longer compilation time. Based on this added flexibility in compilation, this work proposes a multilevel (multi-layer) graph partitioning framework for solving the mapping and scheduling problem for distributed heterogeneous embedded systems. However, this work concentrates on developing a multi-layer task graph clustering scheme as a part of a multi-step scheduling approach. This work introduces linear clustering techniques for vertex matching in Directed Acyclic Graphs (DAGs) that enable the construction of multi-layer coarsened DAGs with different granularity levels. This paper addresses the major obstacle facing the DAG contraction process, which is the formation of cycles, and discusses this issue for direct and chained dependence cycles. Solutions are provided that establish the basis for performing cycle testing in order to avoid the formation of cycles under certain conditions. This research shows, by analysis, that the upper-bound time complexity of the proposed cycle testing method is O(|V|3), where V is the set of task nodes. Also, this research shows, by analysis, that the complexity of the proposed multi-layer DAG coarsening scheme is O(|V|6). The simulation results verify the cost-effectiveness of the proposed scheme, and show its capabilities in generating clusters at different granularities.
KW - Distributed real-time systems
KW - Embedded multiprocessor systems
KW - Inter-process communication
KW - Mapping and scheduling
KW - Multilevel graph partitioning
KW - Task clustering
UR - http://www.scopus.com/inward/record.url?scp=84866548041&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84866548041&partnerID=8YFLogxK
M3 - Article
AN - SCOPUS:84866548041
SN - 1076-5204
VL - 19
SP - 187
EP - 202
JO - International Journal of Computers and their Applications
JF - International Journal of Computers and their Applications
IS - 3
ER -