Wow,

I tried your "somewhat ugly" workaround for windows -
ie I took out the socketerror() test if connect()
returns zero and it runs perfectly. Very quick, and
no CPU maxout, even with 400 simultaneous hits.

Do you think this is a reasonable workaround? The
MSDN link you sent seems quite clear-cut : a 0
returned from connect() indicates the connection was
made.

FYI here's the libcurl code I modified for this test
(starts line 685 in connect.c, libcurl 7.10.7) :-

    if(0 == rc) {
#ifdef WIN32
                // win32 : presume 0 means truly connected
                *connected = TRUE;
                break;
#else
      int err = socketerror(sockfd);
      if ((0 == err) || (EISCONN == err)) {
        /* we are connected, awesome! */
        *connected = TRUE; /* this is a true connect
*/
        break;
      }
      /* nope, not connected for real */
      rc = -1;
#endif // WIN32
    }

Any problems with this? I'll have a look around online
to see if I can pick up any other info.

To answer your other questions, the minimum seems to
be about 50. It's hard to pin down an exact number,
because the behaviour isn't always consistent -
sometimes it was fine with 50 threads. Suggests a
race condition in winsock2 IMHO. The server was
running on the same LAN as the test machine.

Thanks lots for your help so far!

rgds,
Andrew Fuller

--- Daniel Stenberg <daniel_at_haxx.se> wrote: > On Tue,
14 Oct 2003, Andrew Fuller wrote:
>
> > By the way, if it's any use, this time the longest
> path (timewise) according
> > to Rational Quantify was :
> >
> >
>
curl_easy_perform->Curl_perform->Curl_connect->SetupConnection->ConnectPlease->Curl_connecthost->socketerror->getsockopt->EnterCriticalSection
> >
> > Next longest was the same to above except ended :
> >
> > ...->getsockopt->NtRemoveIoCompletion
>
> Thanks for the provided info. Here's my take on
> analyzing it:
>
> The VERBOSE output log (from your previous mail) is
> a mess to read and really
> doesn't help much since all the different threads'
> output mix in a seemingly
> random way. You would need to prefix all output
> lines with a thread id/number
> or something for it to start becoming useful. (Thus
> you'd need to use
> DEBUGFUNCTION.)
>
> How many threads do you need to get this problem?
> You used 80 here, but what's
> the minimum amount? What's your network connection
> to the test server?
>
> We could try to learn from history. See what Chicco
> wrote back in 2002:
> http://curl.haxx.se/mail/lib-2002-02/0057.html
>
> Bjorn Reese filled in some more info on why libcurl
> does this:
> http://curl.haxx.se/mail/lib-2002-02/0058.html
>
> The relevant question here and now is of course: can
> we modify this behavior
> on Windows to make it run better? Chicco said that
> he didn't need to invoke
> 'socketerror' if connect() returns 0. Perhaps this
> is always true on windows?
> This page on microsoft.com seems to claim this:
>
http://msdn.microsoft.com/library/default.asp?url=/library/en-us/winsock/winsock/connect_2.asp
>
> I don't use windows enough myself to try this out.
>
> --
> Daniel Stenberg -- curl: been grokking URLs since
> 1998
>
>
>
-------------------------------------------------------
> This SF.net email is sponsored by: SF.net Giveback
> Program.
> SourceForge.net hosts over 70,000 Open Source
> Projects.
> See the people who have HELPED US provide better
> services:
> Click here: http://sourceforge.net/supporters.php
>

________________________________________________________________________
Want to chat instantly with your online friends? Get the FREE Yahoo!
Messenger http://mail.messenger.yahoo.co.uk

-------------------------------------------------------
This SF.net email is sponsored by: SF.net Giveback Program.
SourceForge.net hosts over 70,000 Open Source Projects.
See the people who have HELPED US provide better services:
Click here: http://sourceforge.net/supporters.php
Received on 2003-10-15