[Beowulf] torus versus (fat) tree topologies
Many of your questions may have already been answered in earlier discussions or in the FAQ. The search results page will indicate current discussions as well as past list serves, articles, and papers.
Patrick Geoffray patrick at myri.comThu Nov 18 20:39:23 PST 2004
- Previous message: [Beowulf] torus versus (fat) tree topologies
- Next message: [Beowulf] torus versus (fat) tree topologies
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Ashley Pittman wrote: > On Fri, 2004-11-12 at 17:51 -0500, Patrick Geoffray wrote: > >>Going trough a crossbar cost about 100-150ns these days. > > > This isn't necessarily true, we reckon about 25ns on elan4. This > doesn't give you perfect latency scaling across very large networks but > it's very close. It's ~50ns on the Myrinet Xbar32 but you need to add the one-time cost of the SerDes for the fiber (~100 ns). You can do without if you use bulky noisy degrade-with-time copper cables :-) ). Of course, the crossbar overheads add up and the number of crossbars to cross depends on the number of ports per crossbar. Last time I looked, the crossbars on the Quadrics Elite switch had 8 ports each, but it may be more by now. Patrick -- Patrick Geoffray Myricom, Inc. http://www.myri.com
- Previous message: [Beowulf] torus versus (fat) tree topologies
- Next message: [Beowulf] torus versus (fat) tree topologies
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
More information about the Beowulf mailing list
