On Mar 19, 2006, at 11:05, Nicholas Clark wrote:
Is the choice of implementation actually visible to a user of the API?
If "yes", where, and can we avoid it?
No.
At which point we have flexibility in
how things are actually implemented.
Yes.
I can see that emulating asynchronous ops with synchronous ops and
POSIX
threads is portable, whereas a platform specific AIO implementation
with
synchronous ops implemented atop it is likely to offer higher
performance.
The I/O layers implementation allows us to have multiple different
combinations. We could have sync ops + threads for one version, sync
ops on top of system AIO ops for another, sync ops as system sync ops
+ async ops as system AIO ops for another. (My guess is that using
the system sync ops for the synchronous versions will actually get
better performance than wrapping synchronous versions around system
AIO ops. Though, of course, the biggest hit in any I/O implementation
is the I/O itself.)
If the API is strong enough, would it let us just write the first
implementation, and allow people to write platform specific
optimisations
later?
Yes.
I would probably start with a sync + threads implementation, and a
layer for Linux that uses the system AIO ops for the async ops
(leaving the sync ops unchanged). That would give us portability and
a proof-of-concept on customization.
Allison
