TY - GEN
T1 - Adaptive strategy for one-sided communication in MPICH2
AU - Zhao, Xin
AU - Santhanaraman, Gopalakrishnan
AU - Gropp, William
PY - 2012
Y1 - 2012
N2 - The one-sided communication model supported by MPI-2 can be more convenient to use than the regular two-sided communication model and has potential to provide better performance. The MPI-2 standard gives flexibility about when RMA operations can be issued and completed. The current MPICH2 implementation employs a lazy approach, in which operations are queued up and issued in the later synchronization phase. This has certain benefits for small data transfers because of reduced network operations, but for large data transfers, issuing operations in an eager fashion could achieve better performance. In this paper we describe our design and implementation of an adaptive strategy for one-sided operations and synchronization mechanisms (fence, post-start-complete-wait, lock-unlock) supported by MPI-2, which combines benefits from both lazy and eager approaches. Our performance results demonstrate that our approach performs as well as the lazy approach for small data transfers and achieves similar performance as the eager approach for large data transfers. In addition, it achieves good overlap of communication with computation.
AB - The one-sided communication model supported by MPI-2 can be more convenient to use than the regular two-sided communication model and has potential to provide better performance. The MPI-2 standard gives flexibility about when RMA operations can be issued and completed. The current MPICH2 implementation employs a lazy approach, in which operations are queued up and issued in the later synchronization phase. This has certain benefits for small data transfers because of reduced network operations, but for large data transfers, issuing operations in an eager fashion could achieve better performance. In this paper we describe our design and implementation of an adaptive strategy for one-sided operations and synchronization mechanisms (fence, post-start-complete-wait, lock-unlock) supported by MPI-2, which combines benefits from both lazy and eager approaches. Our performance results demonstrate that our approach performs as well as the lazy approach for small data transfers and achieves similar performance as the eager approach for large data transfers. In addition, it achieves good overlap of communication with computation.
KW - MPI implementation
KW - MPI-2
KW - One-sided communication
KW - adaptive strategy
UR - http://www.scopus.com/inward/record.url?scp=84867648742&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84867648742&partnerID=8YFLogxK
U2 - 10.1007/978-3-642-33518-1_7
DO - 10.1007/978-3-642-33518-1_7
M3 - Conference contribution
AN - SCOPUS:84867648742
SN - 9783642335174
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 16
EP - 26
BT - Recent Advances in the Message Passing Interface - 19th European MPI Users' Group Meeting, EuroMPI 2012, Proceedings
T2 - 19th European MPI Users' Group Meeting on Recent Advances in the Message Passing Interface, EuroMPI 2012
Y2 - 23 September 2012 through 26 September 2012
ER -