sorry ... not cut through

Richard Walsh rbw at ahpcrc.org
Wed May 15 11:56:23 PDT 2002


On Tue, May 14, 2002 at 12:18:56PM Greg Lindahl wrote: 

> Cut-through vs. store-and-forward ought not affect bandwidth at all.

Assuming both types of switches fully pipeline successive forwards
and/or the cut-through vs. store-and-forward latency differences are
a small percentage of the whole message/packet transfer time ... Yes?

Can we assume the former for all swithes? Clearly they should be designed
that way, but we did not get pre-fetch piplines from memory until fairly 
recently.  Does someone know? The switch type might affect observable 
bandwidth for small to medium size messages where it is smaller to begin 
with especially if pipelining is absent.

In the limit of small un-pipelined messages, latency defines bandwidth.

rbw




More information about the Beowulf mailing list