[JDEV] Keep-Alives

[Matthias Wimmer]

Hi Tijl!

Tijl Houtbeckers wrote:

>A few problems I see with using keep-alives (as far as I understand them) 
>1. It is very low level, not all programming languages / devices will have acces to it.
>
It's right that not every programming language support activating 
keep-alives, but:
- you only have to activate it at the server
- you don't have to do anything on the client it's completely 
transparent to it
- it's accessible by much more programming languages then the OOB data

>2. It is bound to TCP/IP, I could for example use a bluetooth link or a serial link instead 
>of TCP/IP. Jabber over HTTP also sort of falls into this catagory.
>
TCP keep-alives aren't and won't be needed by the Jabber protocol. It 
just solves some problems we have with TCP. Therefore we only need it 
with TCP.

>3. keep alive is OS dependant. Not all OSes provide acces to it, and some let you set 
>it only for all applications, not specific ones (as I understand from earlyer postings).
>
On the one hand I know no operating system that doesn't support it, on 
the other hand you only need support for it on the server side.
Technical keep-alives is just the retransmission of already send IP 
packets. The IP stack of the server that receives this packet will just 
think that this is a packet that he already got but the sender didn't 
get its acknowledge. Therefore it sends the acknowledge again ... the 
server IP stack knows that the cient is still there.
This meens:
- The IP packets are just normal, they can't be blocked by any firewall 
or intrusion detection system
- it's the smalles amout of data we can transmit over an IP network. 
Because the sender knows that the packet is already acknowleged and the 
client will just drop the packet and acknowledge again it can send any 
content it wants in the packet. Therefore most implementations (Linux 
included) send emtpy packets.

Again OOB is the complete difference: It has to be supported by both 
sides of the connection. Even worse: Windows 95, Windows 98 and Windows 
NT clients will crash if they havn't an installed bugfix to their socket 
code and get OOP data. Therefore (even if the server sends the data only 
if the client tells it does support it) the computer of the user will 
crash if he has an broken OS ... and he won't start the client anymore.

>4 I asume keep-alives will be for *all* user that log into jabber, some might not want 
>keep-alives or anything like it at all for their clients.
>
As I told: It's completely transparent to the client. The only thing why 
he couldn't like it is that the client has very expensive metered 
traffic. For this clients would be possible to include some way to 
disable keep-alives (the server can activate and deactivate keep-alives 
per connection). But as I wrote above: It is a very small amount of 
additional traffic and it's only generated on idle connections. Every 
other way to fix our TCP problems will generate more traffic.

>A few of these problems came up before on the mailinglist, and I haven't yet heard a 
>solution for them. It might not hinder any implemention for this on *nix, but the 
>problem will still remain for the platforms/implementations that can not use this.
>
We have servers on Unix, Windows and Java. All of this three platforms 
support keep-alives (Java since version 1.3). The client side is 
supported by every platform/language that supports TCP/IP.


Tot kijk
   Matthias

-- 
Fon: +49-700 77007770		http://matthias-wimmer.de/
Fax: +49-89 312 88654		jabber://mawis@charente.de