async disk io

On Wed, Apr 05, 2000, Adrian Chadd wrote:
>
> My current pass through the modio code is to prepare the storage
> manager for some sane memory efficiency (memory sizes and copies)
> so an async disk IO model fits in nicer.
>
> Is anyone running squid with larger SM_PAGE_SIZE / DISK_PAGE_SIZE
> values? I'm also looking at killing DISK_PAGE_SIZE totally to
> reduce the number of memory copies when swapping out.
>
> Ideas/opinions ? And no, I'm not touching the HTTP code to improve
> its efficiency. :-)

Ok, obviously noone is doing this. :-) So, I ripped out DISK_PAGE_SIZE
totally, and now for swapouts it walks the mem_node list and writes
SM_PAGE_SIZE or less (the code makes sure that if the object is still
STORE_PENDING there is at least SM_PAGE_SIZE to swap out, so it
shouldn't write partially filled buffers unless they're not going to be
appended to. I hope :-) Today I'm going to be working on swapin and
store_client to bring some more copy-sanity there.

Now, I want to finally move onto implementing async disk IO. So,
the question is: what mechanism do people think I should use for
this? I have my own clever ideas, but I'm very FreeBSD/Solaris
centric ..

Adrian
Received on Fri Apr 07 2000 - 06:14:03 MDT