Re: Re: [PATCH 3/3] add `#kgio_writev` and `#kgio_trywritev`

[Eric Wong <normalperson@yhbt.net>]

Юрий Соколов <funny.falcon@gmail.com> wrote:
> 2012/5/31 Eric Wong <normalperson@yhbt.net>
> 
> > Юрий Соколов <funny.falcon@gmail.com> wrote:
> > > With many tiny strings substitution function from writev_compat.h is
> > faster.
> > > It seems that we should not always make call to system writev, but do it
> > > depending on a data shape. I'll try play with tomorrow.
> >
> > Yes, I expect writev_compat (or even naive write(ary.join)) to be faster
> > for tiny strings.  Each iovec struct is 16-bytes on 64-bit, and this is
> > a _huge_ overhead on small strings.  Setting up pointers/lengths ends up
> > being more expensive than copying the strings themselves.
> >
> > I think this optimization is best left to application authors to
> > decide (and documented as such).
> >
> 
> Well, it is faster even with prepared iovec , and joining string on Ruby
> level is still
> slower.
> I think, it is easy to decide automatically whether to choose custom
> function or
> call library function considering average string size in a batch:
> when average string size is about 1000 bytes library function is certainly
> faster
> when strings are of 100 bytes then custom function is much faster
> Taking something around 500~700 bytes as threshold we could have a good
> performance in every case.

OK, numbers sound reasonable.  Do you have benchmarks/examples for this?

> I don't think 16 bytes is very huge overhead considering that buffer is
> reused
> for every batch of strings.

I meant userspace needs to copy 16 bytes of metadata into the kernel
for every iovec.