Re: Looking for Linux equivalent of IO Completion Ports...

From: QNils_O=2E_Sel=E5sdal=22?= (NOS_at_Utel.no)
Date: 08/25/05 

Date: Thu, 25 Aug 2005 09:37:10 +0200

David Schwartz wrote:
> ""Nils O. Selåsdal"" <NOS@Utel.no> wrote in message
> news:430c3638@news.broadpark.no...
>
>>David Schwartz wrote:
>>
>>>"WTH" <spamsucks@Ih8it.com> wrote in message
>>>news:IEkNe.23453$Rm3.694@bignews4.bellsouth.net...
>>>
>>>
>>>
>>>>You're totally missing the point. You can already avoid thread thrashing
>>>>with asynchronous I/O in Linux but the methods I'm aware of (poll/select)
>>>>do not scale well at all because they take literal FD by FD scans by the
>>>>kernel to discern changes. When you get into the 'thousands' range, the
>>>>time for these scans becomes significant.
>>>
>>>
>>> They scale perfectly. The idea that poll does not scale well is a
>>>complete myth.
>
>
>>Not it's not.
>>Take a look at libevent. (please ?)
>>Take a look at why epoll and /dev/poll was invented.
>>Take a look at the kernel side of these implementations.
>
>
> You snipped my justification, which I will repeat:
>
> In fact, under most realistic conditions, 'poll' gets more efficient as
> the number of connections goes up because the greater work to be done after
> each call to 'poll' means more time before the next call to 'poll', which
> means a greater percentage of the sockets probed will be discovered to be
> ready.
>
> If you think this is incorrect, you are welcome to try to refute it.
Ok, sorry about that. No, you're quite correct in in those statements.
You also mention in another post that you divide up your poll sets in
1000 connections per thread, which is fine, and quite a clever way
to achieve scalability.
My point was you have to do something like this, as having poll much
higher doesn't scale :-) 

-- 
Nils O. Selåsdal
www.utelsystems.com

Relevant Pages

  • Re: POLL -- Violence Against Women
    ... "bed" in the poll question would cause ANYONE ... any of that as any kind of a justification for the man to ... possible that a person who punches a slot machine after ...
    (misc.writing)
  • Re: Are a set and trips different ?
    ... set involves a pair in the hole. ... I can't believe you started a poll to "prove your point" LOL, ... Board" and they just repeat it ad nauseum. ...
    (rec.gambling.poker)
  • Re: McMisanthrope Top 25: January 11, 2011
    ... My justification for this poll is that, in the final poll, the only ... thing that really matters is "Which team is the Champion?" ... And, since that team is obviously Auburn, Auburn should get all the ...
    (rec.sport.football.college)