A Performance Model of Many-to-One Collective Communications for Parallel Computing

Author

Lastovetsky, Alexey ; O´Flynn, Maureen

fYear

2007

fDate

26-30 March 2007

Firstpage

1

Lastpage

8

Abstract

This paper presents a performance model of many-to-one collective communications for MPI platforms on a switched Ethernet network. The model is based on empirical findings from observation of many-to-one operations over a wide range of message sizes. The model reflects a significant increase in the execution time for medium-sized messages, persistently observed for different parallel platforms and MPI implementations and not reflected in traditional communication performance models. We also demonstrate that the use of the model can significantly improve the performance of parallel applications, intensively using many-to-one communications.

Keywords

local area networks; message passing; parallel processing; MPI platform; many-to-one collective communication; message passing; parallel computing; performance model; switched Ethernet network; Bandwidth; Communication switching; Concurrent computing; Delay; Ethernet networks; Frequency; High performance computing; Parallel processing; Probability; Wide area networks; Cluster computing; Collective communications; Communication networks; High performance computing; MPI; Many-to-One communications; Parallel computing; Performance modeling;

fLanguage

English

Publisher

ieee

Conference_Titel

Parallel and Distributed Processing Symposium, 2007. IPDPS 2007. IEEE International

Conference_Location

Long Beach, CA

Print_ISBN

1-4244-0910-1

Electronic_ISBN

1-4244-0910-1

Type

conf

DOI

10.1109/IPDPS.2007.370574

Filename

4228302