To determine disconnects within the second timeframe, that means you'd have to
configure TCP keep-alives to be occurring nearly once a second. Since the
keep-alives may be bi-directional, assuming one per second, you'd be sending
about 2 * 54 bytes * 60 * 60 * 24 * 30 = 267 MB of data a month per connection.
I hope you don't have too many clients, and that the clients ("mobile device")
don't have bandwidth limited data plans. ;-)
I've also read a bit that cellular carriers can sometimes decide to drop TCP
level keepalives. So, with a TCP timeout window in the second range,
connections may end up being more flakey than they really are. The cellular
carriers dropping TCP keepalives is one of the reasons we use application level
keepalives (in addition to being able to configure keepalives on a per
connection basis).
Chris
________________________________
From: Zigor Salvador <[email protected]>
To: "[email protected]" <[email protected]>
Sent: Wed, October 13, 2010 1:49:18 AM
Subject: Re: almost immediate detection of disconnected client
Let me resurrect this old thread to point to this URL:
http://www.gnugk.org/keepalive.html
I managed to change the keepalive timer of my system so that I can detect
(IOException) the disconnection of, say, a mobile device within seconds of its
disconnection.
It's far from ideal, since this configuration affects all of the system, but
this way you don't have to implement a heartbeat protocol to detect "almost
immediate" disconnection.
Zigor.
On 7 Jul 2010, at 16:55, Alexander Christian wrote:
>
> Hi,
>
>
> On Wed, 07 Jul 2010 14:16:25 +0200, Emmanuel Lecharny
> <[email protected]>
> wrote:
>> There is no way to detect a connection failure. The connection is opened
>
>> or closed, but you will know only after having successfully sent
>> something to the client.
>
> What about TCP KeepAlive? That IS checking the connection automatically.
> The only problem with this is:
> You don't know WHEN the OS will send a keep-alive-message to check the
> connection. It can be within seconds (unlikely), several minutes (more
> likely) or even hours ... (or maybe never?)
> It is possible to change the keep-alive interval, but this then counts for
> all socket connections the OS manages ... :-(
>
> A maybe useful hint:
>
> http://mindprod.com/jgloss/socket.html#DISCONNECT
>
> I don't use transactions. But I also wanted to know asap when the
> connection dies.
>
> So I created my own keepl-alive check:
>
> * As long as client and server and sending data to each other, the check
> happens when sending this data. If sending fails, the connection is broken
> * If client and server fall into IDLE (mina's session idle thing), then my
> communicationlayer on top of MINA sends ping-pong packages: If a pong has
> not been received xx sec after a ping is sent, the connection is broken.
>
> For me this has the positive sideeffect, that a "blocking" remote
> (dealocked?!) is detected as well.
>
> br,
> Alex
