[OpenAFS-devel] how does fileserver read from disk?

[Steve Simmons]

On Tue, Sep 20, 2005 at 06:59:20AM -0400, Tom Keiser responded to
Marcus Watts:

> 6) even for those platforms that are stuck in the early 90's using
> pthreads to emulate kernel aio, there is a distinct advantage . . .

IMHO the advantage on those older box is not relevant (which is not
to say that there isn't an advantage, quite the opposite).  As best I
can tell, neither disk throughput nor network throughput are anywhere
near as good as they should be for AFS servers.  If FooNIX does a bad
job of implementing aio, *that's not our problem.*  It's the problem of
the folks who are using FooNIX for AFS servers.  And it's probably not
even a problem for them.  If an aio-based AFS implementation doesn't
give them a win, all they've missed is a win.  At worst case, they can
stick with current implementation.

Or to be more blunt about it, if a few old boxes lose while everyone
else wins big, then we should go for it.  The folks with those old
boxes can stick with their existing hw/sw combos with no lossage, and
if they want the wins, they need to upgrade their file servers.  If
they cannot do so, they've got bigger problems than worrying about
getting another 10% in throughput out of their old servers.

And it's not like any of this is going to spring into existence tomorrow;
1.2 was released four years ago last Monday (yes, I know there were
compelling reasons).  God help us if 2.0 or 1.6 takes another four years,
but two wouldn't surprise me a bit.  Whatever it is, even fewer of those
legacy boxes will be around by then.
-- 
Sturgeons Law:      "90% of everything is crap."
Simmons' Corollary: "10% of everything is worth looking at."