Re: async I/O seems to be blocking on 2.6.15

[Jens Axboe <jens.axboe@oracle.com>]

On Mon, Nov 06 2006, Brent Baccala wrote:
> On Sun, 5 Nov 2006, Jens Axboe wrote:
> 
> >On Fri, Nov 03 2006, Brent Baccala wrote:
> >>
> >>Does 7 microseconds seem a bit excessive for an io_submit (and a
> >>gettimeofday)?
> >
> >I guess you mean miliseconds, not microseconds. 7 miliseconds seems way
> >too long. I repeated your test here, and the 100 submits take 97000
> >microseconds here - or 97 miliseconds. So that's a little less than 1
> >msec per io_submit. Still pretty big. You can experiment with oprofile
> >to profile where the kernel spends its time in that period.
> >
> >-- 
> >Jens Axboe
> >
> 
> Yes, of course, milliseconds.  I have enough other problems with this
> program (measured in minutes, and no mistake that) that I doubt I'll
> be profiling the kernel any time soon, but thank you for your help.
> 
> More than anything else, you've made me understand that I can't just
> fire off a bunch of async requests like I'm tossing peanuts across the
> table.  I've really got to pay attention to what's in that kernel
> queue and how it gets managed.

Yeah, I'm afraid so. We really should be returning EAGAIN or something
like that for the congested condition, though.

-- 
Jens Axboe