On Tue, 2010-08-24 at 16:30 +0200, Mladen Turk wrote:
> Work in progress ...
> Will be used in new Poll.interrupt API so we can remove all
> that clumsy pollTime/maintainTime logic for adding new sockets
> to the pollset.
> It should lower the accept latency, remove the need for
> socket add queue and remove unnecessary thread context switches.

What is the usefulness of constantly interrupting polling ? Surely this
will have a cost too. Right now, while polling is in progress, sockets
are queued, and will be added all at once, which presumably will limit
the amount of thread interactions that would require syncing (and only
adding to the queue is synced, which could actually be optimized away by
using some sort of concurrent structure, but unfortunately this is not
so easy to write). Also, maintain does timeouts, which are not done by
polling, I don't see how this can be dropped.

So I don't understand how this can perform better, but feel free to
experiment ;) Maybe on Windows it would yield something (without any
accept filter/defer/whatever, it will suffer from the ms-scale latency,
which is normally not a real world issue, but if it wasn't there it
would be nicer), at the expense of the decent platforms.

Rémy



---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@tomcat.apache.org
For additional commands, e-mail: dev-h...@tomcat.apache.org

Reply via email to