From: Jason Duell (jcduell_at_lbl_dot_gov)
Date: Fri Apr 28 2006 - 12:50:35 PDT
On Fri, Apr 28, 2006 at 02:17:45PM -0500, Jason Beech-Brandt wrote: > Hi, > > Does anybody have experience using the "symmetric" shared pointer > representation in Berkeley upc? I've built a version of upc-2.2.2 for > an SGI Altix with the --enable-sptr-symmetric configure option and then > built an application code (shmem network) which contains a large number > of shared pointers with indefinite blocksize. The results I'm getting > when using the symmetric/packed representation are consistently slower > than just using the packed representation. The problem gets worse the > larger the number of threads I use. > > Does anybody have any tips for the recommended use of the symmetric > representation? I understand they are still experimental, I just want > to make sure I'm not using them in an unintended fashion... We're in the process of tuning the symmetric pointer representation. Since they use native pointers under the covers, they should eventually be faster than the packed pointers. But I know of at least one bug which I just tracked down that slows down their performance somewhat. Generally, you're usually better off casting from shared to to regular C pointers if possible before doing a series of "*p++"-type operations. But by next release we'll hopefully have symmetric running at or near native pointer speeds, at least for some platforms (for others there will always be at least the cost of performing an 'if' test to see if remote network calls are needed). -- Jason Duell Future Technologies Group <jcduell_at_lbl_dot_gov> Computational Research Division Tel: +1-510-495-2354 Lawrence Berkeley National Laboratory