Collective communication costs analysis over gigabit ethernet and infiniband

Authors:
Hyacinthe Nzigou Mamadou;Takeshi Nanri;Kazuaki Murakami
Affiliations:
Department of Informatics, Kyushu University, Kasuga-shi, Japan;Computing and Communications Center, Kyushu University, Fukuoka-shi, Japan;Department of Informatics, Kyushu University, Kasuga-shi, Japan
Venue:
HiPC'06 Proceedings of the 13th international conference on High Performance Computing
Year:
2006

Citing 6
Cited 0

LogP: towards a realistic model of parallel computation

PPOPP '93 Proceedings of the fourth ACM SIGPLAN symposium on Principles and practice of parallel programming
The communication challenge for MPP: Intel Paragon and Meiko CS-2

Parallel Computing
LogGP: incorporating long messages into the LogP model—one step closer towards a realistic model for parallel computation

Proceedings of the seventh annual ACM symposium on Parallel algorithms and architectures
Automatically tuned collective communications

Proceedings of the 2000 ACM/IEEE conference on Supercomputing
Fast Measurement of LogP Parameters for Message Passing Platforms

IPDPS '00 Proceedings of the 15 IPDPS 2000 Workshops on Parallel and Distributed Processing
Performance Analysis of MPI Collective Operations

IPDPS '05 Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS'05) - Workshop 15 - Volume 16

Quantified Score

Hi-index	0.01

Visualization

Abstract

Users of parallel machines need to have a good grasp for how different communication patterns and styles affect the performance of message-passing applications. MPI Collective communications involve multiple processors, and their performance prediction is a tricky task to perform. In order to evaluate the performance of collective communications, we attempt to extend LogGP and P-LogP standard point-to-point models. Our objective is to compare these models with the empirical data, and identify the most suitable for performance characterization of collective communications. The models proposed are related with the implemented algorithms in MPICH. The experimental results performed on clusters of 16 and 64 processors connected by Infiniband and Gigabit Ethernet networks respectively, show the same trend. For any collective operation, given a number of processors and a range of message sizes, there is at least one model that predicts the performance precisely.