On 12/21/19 11:45 AM, Benjamin Mahler wrote:
> + Jens
> 
> Sounds like some of the iouring findings are surprising to Jens (the
> author).
> 
> Is there a benchmark he can run to look into this?
> 
> Do you have more explanation about "silently ignore parts of the
> requested events on an undocumented subset of file description types"?

Switch to private email, it's more convenient. I just sent this one to
Marc separately, FWIW:

Someone pointed me at this:

https://www.mail-archive.com/[email protected]/msg02767.html

And while I'm glad to see some io_uring adoption, I'm a bit puzzled
at some of this. First of all, I downloaded the release and looked at
your 'general notes' about io_uring. Comments on some of them:

e) what kernel are you using? 5.3 (iirc) and newer supports two mmaps
   for this, mapping the two rings in one call. the sqes are separate,
   so can't really be folded with this.
g) but you have timeout commands, if you want to get woken.
h+i) 32k is max today, and since 5.5 we handle overflows much more
   gracefully. io_uring never drops events anymore.
j) where are the details on this? you mention this in the release as
   well, but this is the first I've heard of this. aio and io_uring
   use the same mechanism for this, so not surprising that they share
   whatever quirk this is. doesn't mean it can't get fixed up though,
   but hard to fix something up if nobody told me about it.

For the posting, I'd love to be able to benchmark this. How do you
benchmark against libev using different backends?

And back to e), what kernel did you use? We're obviously maturing
this API (and adding features) as we go, and while most bugs fixes
do make it back to -stable, that's not always true. I'd love to know
more about oopses you've seen, and on what kernel you ran.

-- 
Jens Axboe


-- 
Jens Axboe


_______________________________________________
libev mailing list
[email protected]
http://lists.schmorp.de/mailman/listinfo/libev

Reply via email to