Re: async-io

On Sun, Apr 23, 2000, Henrik Nordstrom wrote:
> Adrian Chadd wrote:
>
> > Under FreeBSD it'd probably work better to use the posix AIO or
> > the kqueue stuff, under Linux it'd probably work better to
> > useclone()/threads, under Solaris libaio might bethe best choice ..
> > Implementing disk as the underlying mechanism forces me to clean up
> > the storage manager and its use of memory first, which isn't a bad
> > thing.
>
> Some comments:
>
> Squids async-io is implemented because Posix AIO cannot do what Squid
> requires, and also because Solaris AIO implementation at the time did
> not manage to keep up with it in a stable manner.
>
> One major shortcoming of Posix AIO is that it doesn't define AIO
> operations for open/close. Only read/write/cancel/sync are defined. The
> other major shortcoming is in how it signals completetion.

libaio is much nicer at doing that, but I do believe it doesn't define
open/close. That does limit its usefulness. Good point.

>
> I think the FreeBSD implementation has extended AIO with open/close
> calls, but if not then you probably want to use some other alternative
> like diskd.
>
> async-io (or aiops.c) is a Posix threads based AIO like implementaion
> implementing open/close/read/write/cancel/stat/unlink/sync. On top of
> that there is a callback handling layer (async_io.c).

Having a modified aiops.c which had the concept of a 'queue' would be
nice. I ight look into this at a much later date.

> Since aiops.c isn't really AIO the two layers should probably be merged
> into one layer. The reasons why it is separated in two is that Stewart
> wanted to have a "userspace" vs "kernel" separation where data is copied
> to protect himself from aborts changing the data under the feets of an
> ongoing operations, but as you say you are working with fixing this
> issue of the store magager it shouldn't be that much of an issue.

Well, I'm not sure what has change sicne Stewarts aiops implementation,
but the storage manager still left the memry in the MemObject intact
while the copy was being written out.

Adrian
Received on Sun Apr 23 2000 - 10:58:08 MDT