On Wed, Mar 08, 2006 at 04:14:14PM -0800, Matthew Dillon wrote: > It would be disaster. It would become extremely difficult to track down > bugs. And because of that I think its an ultra bad idea. It is not > worth the ~1 uS savings per system call. It just isn't.
Esp. since I don't believe the actual number of system calls to go down. What is better in this approach is buffering for connections with high bandwidth delay product (I can't get more than 10 MByte/s for non-local TCP links, even though the link is far from being saturated, since parallel TCP connections are fine). Another important aspect is the processing latency is not necessarily going down as well. As a matter of fact, something like having a great number of mostly CPU bound daemons pushing out the results of the calculation (think PHP) can now result blocked transmission for the duration of a full time slice. That's something the kernel normally handles. Joerg
