On Sun, Apr 23, 2000, Henrik Nordstrom wrote:
> Adrian Chadd wrote:
>
> > 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.
>
> ???
>
> aiops.c has a queue, and a round-robin scheduler that shedules entries
> from the queue on individual threads as they become available.
>
> What should perhaps be changed in that area is that the threads should
> probably pick entries from the queue themselves and deliver them to a
> single (or small pool of) return queues when done. Stewart was perhaps a
> bit too worried about mutex congestion, so currently there is a mutex
> per thread rather than for the queue(s).
I meant a queue of operations which is maintained linearly. Then you
can look at implementing elevator sorts and simple clustering/prefetching
where applicable. I don't want to have too much magic in the aiops
code, but enough to prevent replication of code.
Again, I haven't reached a point where I can implement the async IO.
When I get there, I'm sure I'll hit problems I haven't forseen.
Adrian
Received on Sun Apr 23 2000 - 16:57:46 MDT
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:12:23 MST