The one-sided communication model (or remote memory access) supported by MPI-2 is more convenient to use than the regular two-sided send/receive communication model, because it allows the sender to specify all data transfer parameters and avoids the receiver to be explicitly involved in data receiving. One-sided communication model also has potential to provide higher performance. The MPI-2 standard provides flexibility about when RMA operations can be issued and completed, which makes the MPI implementation possible to be optimized internally. The current MPICH2 implementation uses a lazy approach to issue one-sided operations by queuing the operations and issuing them during the later synchronization phase. This has certain benefits with respect to short operations in terms of reduced network operations. For large data transfers, issuing operations in an eager fashion could be more beneficial as well as provide more scope for overlapping communication and computation. In this thesis we describe the design and implementation of an adaptive approach for all three synchronization mechanisms defined in MPI: fence, post-start-complete-wait, and lock-unlock. We evaluate our implementation with both micro benchmarks and the Graph500 benchmark to demonstrate the performance impact of our approach. The performance results show that our hybrid approach performs as good as the lazy approach for small data transfers and achieves performance similar to the eager approach for large data transfers, and its overlapping percentage is good.
【 预 览 】
附件列表
Files
Size
Format
View
Adaptive strategy for one-sided communication in MPICH2
PDF
download
878987797
028-85220240
