The documents distributed by this server have been provided by the contributing authors as a means to ensure timely dissemination of scholarly and technical work on a noncommercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, notwithstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.
Publications of SPCL
|T. Hoefler, C. Viertel, T. Mehlan, F. Mietke, W. Rehm:|
|Assessing Single-Message and Multi-Node Communication Performance of InfiniBand|
(Vol , Nr. , In Proceedings of IEEE International Conference on Parallel Computing in Electrical Engineering (PARELEC'06), presented in Bialystok, Poland, pages 227-232, IEEE Computer Society, ISSN: , ISBN: 0-7695-2554-7, Sep. 2006)
AbstractWe present a micro benchmark suite to evaluate InfiniBandTM implementations with regards to single message performance and the addressing of many hosts. We use a 1:n communication pattern to assess the latency and bandwidth for all different combinations of InfiniBandsTM transport services and functions. The results gathered in this study are used to optimize MPI collective communication operations where 1:n communication schemes are not used widely today. We show that applications as well as collective algorithms can benefit from sending multiple messages in a single round. Moreover, the results will be used to choose the transport service and function to develop InfiniBandTM optimized collective communication functions. Our study compares all available transport options and shows that single packet sends can be very expensive compared to multi packet sends.