From mboxrd@z Thu Jan  1 00:00:00 1970
Message-ID: <96d88ee6d8e687cb0e74c66a90247a62@quanstro.net>
From: erik quanstrom <quanstro@quanstro.net>
Date: Mon,  8 May 2006 14:41:04 -0500
To: 9fans@cse.psu.edu
Subject: Re: [9fans] kernel development: how not to
MIME-Version: 1.0
Content-Type: text/plain; charset="US-ASCII"
Content-Transfer-Encoding: 7bit
Topicbox-Message-UUID: 4f5db5e6-ead1-11e9-9d60-3106f5b1d025

i agree with your point, but i think the problem with splice and tee is 
worse than that.

before complicating the system with this optimization, shouldn't a
real system doing real work (i.e. not benchmarking) be found where
splice and tee could help?

parameters change, the balance between cpu, memory b/w and i/o
changes over time and from machine to machine.

i wonder if splice is really faster than coping the buffer on a 
NUMA machine?

remember: in linux, as in life, it's ready, shoot, aim.

- erik

On Mon May  8 10:44:21 CDT 2006, davel@anvil.com wrote:
> Yawn.
> 
> After the Beserkeley folks implemented vread() and vwrite() in 4.1BSD
> many people (including myself) saw the concept and thought
> "why didn't they just detect that as a special case in read() and write()?":
> exactly the same benefits without the new syscalls.
> 
> In the same way, most of the speed improvements for muck like splice(),
> could probably be gained by tuning the current system
> (and, hey, you might even speed up a couple of other things along the way).
> 
> Of course improving current functionality without complicating any 
> visible interfaces
> is far less fun that implementing the
>     wankmyego(fuckme, im, so, clever, look, at, all, these, parameters, 
> i, must, be, a, genius)
> syscall.
> 
> Sigh,
>     DaveL
> 
> erik quanstrom wrote:
> > sort of.  it's an extension.  here's the kerneltrap summary:
> >
> > 	http://kerneltrap.org/node/6506
> >
> > what linux actually said is that COW games (e.g. for freebsd's zero-copy
> > socket code) are worse than just copying the data.
> >
> > the real difference between bsd's stuff and what linus did is
> > the linux trick is an explict system call, not a vm game.
> > but he misses that adding strange system calls that parallel
> > read and write is even bigger trouble.  either your fundamental
> > object is a file (unix) or a block of memory (multix).  pick a
> > lane!  certainly don't mix the two at the same level.
> >
> > linus may be right that the freebsd vm's techniques are broken.
> > i don't know.
> >
> > the zero-copy idea is tantalizing.  it would be neat to allow the
> > network stacks or devdraw to live in a user-level fileserver
> > without a performance (2 copy) penalty.  but it may be the case
> > that allowing this trickery is more code than it is worth.
> >
> > - erik
> >