Re: [squid-users] Socket option in Squid

On Fri, Aug 17, 2001, Wojciech Puchar wrote:
> > yup. At high connection/data rates this happens.
> > I thought SNDLOWAT set the hard watermark - ie if the incoming
> > data doesn't ever exceed the watermark, it never becomes ready
> > for reading/writing.
>
> and IMHO it will not speedup much. poll will take the same CPU time.
>
> cannot poll table be divided onto few smaller tables depending of
> last-active time and then tables with more active object could be polled
> more often, and others less often?

It'd work, but the real solution is to only register fd's that
we are expecting IO on sometime soon, rather than all of them.

Henrik and I have done some work on this, but the current codebase
needs a little (!) reworking to take advantage of this type of
optimisation.

Adrian
Received on Fri Aug 17 2001 - 13:52:20 MDT