Garrett D'Amore wrote:
bcopy is your friend. :-) Its been this way on SPARC since time
"bcopy is your worst enemy" , there, fixed it for ya! :)
immemorial. The only time DMA binding really made sense was with
*really* slow CPUs (50 MHz sun4m CPUs) or with special hackery (SPARC
dvma special access routines) or with large frames.
Or *really* fast network (like 10GbE) combined with "slow" CPUs
like opterons. The lmbench bcopy bandwidth on most recent hardware
is between about 8Gb/s and 16Gb/s. So not only will bcopy dominate
the CPU, it may also impact the system and reduce memory bandwidth
available for everything else.
There used to be a whitepaper about this which summarized the situation
as "Modern CPUs can bcopy really bloody fast".
The fact that bcopy is 10% of the cost of throughput is *not* a
problem. Once you reduce the the overhead of the stack enough, even
small factors become a problem.
I don't know where you get your 10% figure (100Mb ethernet?), but it
is quite out of date when talking about modern network interfaces,
and should be taken with a truckload of salt when talking about modern
machines with 10GbE or faster interfaces.
Drew
_______________________________________________
networking-discuss mailing list
[email protected]
