RE: Signal driven IO

David Schwartz (davids@webmaster.com)
Tue, 16 Nov 1999 16:00:08 -0800

> Web traffic is very large numbers of slow operations. Your ratio of
> busy/queued is low. I'd agree there should be a point where you
> have a lot of
> busy operations but web/ftp isnt it.
>
> On a T1 you can expect to see 600 parallel connections. Each is averaging
> 2 or 3 K a second with a typical 1500 byte MTU. So each FD wakes
> one or two
> times a second.
>
> So we have 1200 wakeup events per second over 600 handles. Even
> at 100polls/sec
> that means we get 12 positive handles and 588 overheads per poll.
> As it rises
> to the real world values its hits 1 and 599 overhead at 1000 polls/sec
>
> Alan

For 600 parallel connectons with 1,200 wakeup events per second over 600
handles, poll versus select versus signals is irrelevant. Neither of them
will be anywhere near 10% of available computing resources on a P2-450.

I care about 16,000 parallel connections on a switched 100-base-T network.
I care about 'one-to-many' broadcast applications where you receive a stream
from one source and have to propogate that data to 15,000 clients as quickly
as possible. Consider a server propogating streaming audio over the Internet
using TCP connections.

Why optimize for the cases where you aren't under load?

DS

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/