[Beowulf] bandwidth: who needs it?
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.
Greg Lindahl lindahl at pathscale.comThu Oct 21 10:58:55 PDT 2004
- Previous message: [Beowulf] bandwidth: who needs it?
- Next message: [Beowulf] bandwidth: who needs it?
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
> do you have applications that are pushing the limits of MPI bandwidth? > for instance, code that actually comes close to using the 8-900 MB/s > that current high-end interconnect provides? Bandwidth is important not only for huge messages that hit 900 MB/s, but also for medium sized messages. A naive formula for how long it takes to send a message is: T_size = T_0 + size / max_bandwidth For example, for a 4k message with T_0 = 5 usec and either 400 MB/s or 800 MB/s, T_4k_400M = 5 + 4k/400M = 5 + 10 = 15 usec T_4k_800M = 5 + 4k/800M = 5 + 5 = 10 usec A big difference. But you're only getting 266 MB/s and 400 MB/s bandwidth, respectively. Of course performance is usually a bit less than this naive model. But the effect is real, becoming unimportant for packets smaller than ~ 2k in this example. The size at which this effect becomes unimportant depends on T_0 and the bandwidth. -- greg
- Previous message: [Beowulf] bandwidth: who needs it?
- Next message: [Beowulf] bandwidth: who needs it?
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
More information about the Beowulf mailing list
