At 08:59 AM 9/27/00 +0100, Tom Hughes wrote:
>In message <[EMAIL PROTECTED]>
> Dan Sugalski <[EMAIL PROTECTED]> wrote:
>
> > I don't much care how its faked (if it is) as long as it
> > works. Might not be as efficient as full kernel support for async
> > I/O, but it'll do. At least there's some overlap. (You can get
> > better device request ordering and do I/O coalescing if you have
> > more than one request handy)
>
>The problem when I last tried it was that AIO was actually
>less efficient than doing sync IO myself. It might sound
>incredible to somebody used to using VMS async IO like myself
>but it seemed to be true.
>
>Part of it was probably the slowdown that occurred as soon
>as I had to compile -D_REENTRANT and pull in the thread
>support library with its thread safe versions of various
>library routines, and part of it that most of my testing was
>on Unixware where the filing system was vxfs and sync IO goes
>like a train.
Well, in that case we'll not do it on UnixWare. :)
We're probably going to be eating the overhead for threads anyway, but
doing async I/O under the hood will have to be a configurable thing so we
can support platforms that don't support it. Threads themselves should
probably be optional for older platforms, but that's another issue.
Dan
--------------------------------------"it's like this"-------------------
Dan Sugalski even samurai
[EMAIL PROTECTED] have teddy bears and even
teddy bears get drunk
