Marcin Owsiany <marcin@owsiany.pl> wrote:
> On Mon, Jan 07, 2008 at 07:55:43AM +0000, Max Deineko wrote:
>> My first guess would have been that when the server is busy serving an
>> expensive request, the tcp rcv buffer would get full and packets would
>> get dropped.  But wouldn't tcpdump not be able to see them then?

> http://forum.openvz.org/index.php?t=msg&goto=21592
>
> My theory is that when the (receive buffer size / current connections)
> is too small for the window size, then kernel gets itself into a state
> where it's impossible to recover a connection.

Thanks a lot.  I already had found the thread and tried increasing
tcp_rmem size, which hung the webserver completely (maybe it was a litte
too big ;), but I'm still experimenting with the settings.

Of course there might be a problem with other server components, and I'm
not really a networking expert, it's just that the tcp flow doesn't look
like what I'd expect it to.

So is the behaviour I see normal or is there really something wrong at
this level already?  Even if one assumed that userland components were
broken and went haywire sometimes?

Thanks, Max.