On Wed, Mar 17, 2010 at 03:40:13PM +0000, Sad Clouds wrote:
> On Wed, 17 Mar 2010 16:07:04 +0100
> Manuel Bouyer <bou...@antioche.eu.org> wrote:
>
> > On Wed, Mar 17, 2010 at 12:50:51PM +0000, Sad Clouds wrote:
> > > [...]
> > > I profiled a simple event driven server, with 5000 simultaneous
> > > connections. About 80% of the time was spent in write() system call.
> >
> > What makes you think that a syscall writing the same data to multiple
> > descriptors would be faster ?
> > My guess is that the time is not spent in the syscall code itself,
> > but in the socket code (or TCP).
>
> I don't know how much improvement it would make, but I thought it would
> cut down on context switching. Same reason why single writev() performs
> better than multiple write().
The point about writev() is that the write is (more or less) atomic.
This wouldn't be true for a sequence of write() system calls.
Although your code reports a significant time inside write(), I'd suspect
that the system call entry/exit are not responsible.
So merging the code into a single system call would gain very little.
System calls that don't block aren't that expensive on most sane
architectures (ie those that don't need a cache flush to avoid userspace
accessing kernel addresses after the call returns).
David
--
David Laight: da...@l8s.co.uk
