James Carlson writes:
 > The original ARC cases for sendfile and sendfilev (2000/433 and
 > 2000/279) discussed this very issue.  The case owner pointed out that,
 > besides targeting a particular application, the folks implementing it
 > weren't prepared to test all the possible combinations of vnode types
 > to make sure that they'd all work as expected, and thus they wanted to
 > limit the scope as much as possible.

In addition to being flexible, it would be nice if it were able to do
zero-copy more frequently.  I spent quite a lot of time with dtrace
and the OpenSolaris sources trying to figure out what ducks I needed
to have in a row to avoid memory copies in the kernel when I
implemented TCP_SENDFILE support for Solaris in netperf to benchmark
my 10GbE NIC.

As to Linux's UDP sendfile support:  Does anybody know the API? Does
it require the application to make a syscall every MTU-epsilon bytes,
or is there a way to get the kernel to break up the packets?  Whatever
it does, is it actually useful in the digital video environment?

Drew

_______________________________________________
networking-discuss mailing list
[email protected]

Reply via email to