Looking for *nix (esp. Linux) equivalent of IO Completion Ports...

Have read many contradictory statements on the web, I just wanted to know
how to avoid thread thrashing on high performance TCP/IP servers in Linux
(think hundreds of simultaneous active connections [like a web server.])

I'd read something about, in a couple of places, a fairly recent kernel
addition that signals a thread when activity occurs on an FD (but not
something that polls like the bad old days (at several microseconds per
FD...)

Thanks,
WTH

WTH <spamsucks@ih8it.com> wrote:
> Have read many contradictory statements on the web, I just wanted to know
> how to avoid thread thrashing on high performance TCP/IP servers in Linux
> (think hundreds of simultaneous active connections [like a web server.])

Hardware support? I've seen several ads recently in Linux Journal for
a board that appears to claim to do what need. (No I can't remember the
product name, maybe someone else can pitch in?)

Chris

"WTH" <spamsucks@Ih8it.com> wrote in message
news:w4cNe.23284$Rm3.8273@bignews4.bellsouth.net.. .

> Have read many contradictory statements on the web, I just wanted to know
> how to avoid thread thrashing on high performance TCP/IP servers in Linux
> (think hundreds of simultaneous active connections [like a web server.])

Use a thread pool and an "any thread can do any job" architecture. I've
tested up to 16,000 connections with this architecture with no problem.

> I'd read something about, in a couple of places, a fairly recent kernel
> addition that signals a thread when activity occurs on an FD (but not
> something that polls like the bad old days (at several microseconds per
> FD...)

http://www.kegel.com/c10k.html

DS

David Schwartz <davids@webmaster.com> loquated like no one had ever loquated
before with:

> "WTH" <spamsucks@Ih8it.com> wrote in message
> news:w4cNe.23284$Rm3.8273@bignews4.bellsouth.net.. .
>
>> Have read many contradictory statements on the web, I just wanted to
>> know how to avoid thread thrashing on high performance TCP/IP
>> servers in Linux (think hundreds of simultaneous active connections
>> [like a web server.])
>
> Use a thread pool and an "any thread can do any job" architecture.
> I've tested up to 16,000 connections with this architecture with no
> problem.
>> I'd read something about, in a couple of places, a fairly recent
>> kernel addition that signals a thread when activity occurs on an FD
>> (but not something that polls like the bad old days (at several
>> microseconds per FD...)
>
> http://www.kegel.com/c10k.html
>
> DS

Great link, thanks :).

WTH

--
"I notice a former captain of ours said recently that this squad is so
good that we don't need a manager. I took this as a great compliment.
He must have changed his mind since leaving as he said at the time that
Phil Thompson and I would drag the club down. On that point I suppose
he was right - we dragged the club down to Cardiff three times in the
last 10 months." - Gérard Houllier