Re: aio #2: why aren't we doing async writes?

On Wed, Jun 11, 2008, Henrik Nordstrom wrote:

> Simply because when testing on Linux they very rarely block, with the fs
> buffers absorbing them fully as long as you don't run too short on
> memory.

Ah.

> But yes, when you reach the absolute limit of what the drives can
> sustain the buffers will fill up and problems arise, but going async on
> the writes don't really help even then as VM pressure will be too high..

Well, you're assuming all of the IO happens in the syncer, and you're not
blowing the buffer cache in the process.

Considering what side-effect the -log- file writing had when it is done
sync and inline, even when fully buffered, I can't believe that doing the
sync writes will have no effect at high loads.

What I need to do is get this Sun box actually running Solaris and try to
get dtrace to tell me what time is spent in the main thread blocking on
disk writes.

Adrian

-- 
- Xenion - http://www.xenion.com.au/ - VPS Hosting - Commercial Squid Support -
- $25/pm entry-level VPSes w/ capped bandwidth charges available in WA -