From: Paul H. Hargrove (PHHargrove_at_lbl_dot_gov)
Date: Tue Dec 19 2006 - 10:52:30 PST
Konstantin Kleisouris wrote: >Hi everyone, > > Does anyone know what algorithm is used for the implementation of >upc_all_gather? > >Thank you, >Kosta > > Kosta, The current gather algorithm is a "flat tree" topology with all data being sent by the source threads to the destination thread. Depending on the SYNC flags and size of the data, a few different mechanisms may be used to move the data (message-send, RDMA-put, Rendezous-RDMA-put). If you want or need more detail, let me know. The current implementations of the UPC collectives in the Berkeley UPC compiler were written to be correct, but little effort was given to making them the most efficient or most scalable. Work is in progress right now to rewrite the collectives for efficiency and scalability. We hope to offer much improved UPC collectives performance in a future release. -Paul -- Paul H. Hargrove PHHargrove_at_lbl_dot_gov Future Technologies Group HPC Research Department Tel: +1-510-495-2352 Lawrence Berkeley National Laboratory Fax: +1-510-486-6900