You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@river.apache.org by Sim IJskes - QCG <si...@qcg.nl> on 2010/03/08 12:17:20 UTC

time until error

Hello,

does anybody know, or know how to calculate how long it would take 
(taking the complete system into account) from a call to a Remote to an 
error in the following scenario:

A workstation TCP based Endpoint. Service registered. Remote handle 
retrieved by client. Physical network disconnect on server side.

(From the TCP side off things this would constitute complete packetloss 
without other indications (no ICMP for instance, no interface takedown 
at client because of cable disconnect)).

Gr. Sim


-- 
QCG, Software voor het MKB, 071-5890970, http://www.qcg.nl
Quality Consultancy Group b.v., Leiderdorp, Kvk Leiden: 28088397

Re: time until error

Posted by Gregg Wonderly <gr...@wonderly.org>.
Sim IJskes - QCG wrote:
> Hello,
> 
> does anybody know, or know how to calculate how long it would take 
> (taking the complete system into account) from a call to a Remote to an 
> error in the following scenario:
> 
> A workstation TCP based Endpoint. Service registered. Remote handle 
> retrieved by client. Physical network disconnect on server side.
> 
> (From the TCP side off things this would constitute complete packetloss 
> without other indications (no ICMP for instance, no interface takedown 
> at client because of cable disconnect)).

With TCP, keepalive is the only mechanism that can reveal a complete physical 
layer loss for a "reading" socket.  The write retries will eventually produce a 
timeout.  The typical TCP timers for these things are around 3 minutes depending 
on local configuration changes such as with linux where you can "echo" new timer 
values into the kernel /proc tree.

This is another reason why I tend to use smart proxies with leases in them that 
the service has a lease listener on.  In the end, I've found the use of a lease 
normalizes and reveals, the most uniformly, the "lost client" case that needs to 
have cleanup actions performed.  Not only that, but the lease renewal failure in 
the client can cause it to trigger a rediscovery and retrieval of an appropriate 
server instance as well.

Gregg Wonderly


Re: time until error

Posted by Tom Hobbs <tv...@googlemail.com>.
It is a long time.  I've gotten around it in the past by polling the service
which can be done in one of two ways (off the top of my head)...

1)  Add a ping() method to the service and keep hitting that, get a
replacement in the background when it disappears.

2) Lookup the service, get it's ServiceID then keep relooking up the same
service with that ServiceID in the background, replace it when the LUS
doesn't return anything.

Neither of these are perfect solutions.  There's probably another possible
solution; always do lazy lookups just before you need the proxy, then you
can supply your own timeout value, but of course this has the overhead of
always having to do a lookup before each service call.

I find it interesting (and understandable, to a degree) that there doesn't
seem to be any of these kinds of QoS considerations in the spec - that I can
recall...


On Mon, Mar 8, 2010 at 12:04 PM, Sim IJskes - QCG <si...@qcg.nl> wrote:

> Tom Hobbs wrote:
>
>> I don't know where you'd find any official documentation on this, but in
>> my
>> experience the delay has been the TCP timeout.
>>
>> If you already have the proxy then you must wait for the transport
>> layer/protocol to decide that it can't service your request.
>>
>
> Ok, this will mean the attempt will persist up to 9 minutes, solaris 2.2
> cuts it to 2 minutes. (according to Stevens).
>
>
>  I'm not sure about another other versions or OSs.
>>
>
> linux, man 7 tcp, says: 11 to 30 minutes.
>
>
>  Of course, I might be wrong...  :-)
>>
>
> Me too. :-)
>
> That time is too long for failover isn't it?
>
>
> Gr. Sim
>
>
>
> --
> QCG, Software voor het MKB, 071-5890970, http://www.qcg.nl
> Quality Consultancy Group b.v., Leiderdorp, Kvk Leiden: 28088397
>

Re: time until error

Posted by Sim IJskes - QCG <si...@qcg.nl>.
Tom Hobbs wrote:
> I don't know where you'd find any official documentation on this, but in my
> experience the delay has been the TCP timeout.
> 
> If you already have the proxy then you must wait for the transport
> layer/protocol to decide that it can't service your request.

Ok, this will mean the attempt will persist up to 9 minutes, solaris 2.2 
cuts it to 2 minutes. (according to Stevens).

> I'm not sure about another other versions or OSs.

linux, man 7 tcp, says: 11 to 30 minutes.

> Of course, I might be wrong...  :-)

Me too. :-)

That time is too long for failover isn't it?

Gr. Sim



-- 
QCG, Software voor het MKB, 071-5890970, http://www.qcg.nl
Quality Consultancy Group b.v., Leiderdorp, Kvk Leiden: 28088397

Re: time until error

Posted by Tom Hobbs <tv...@googlemail.com>.
I don't know where you'd find any official documentation on this, but in my
experience the delay has been the TCP timeout.

If you already have the proxy then you must wait for the transport
layer/protocol to decide that it can't service your request.

For Solaris 10 you'd probably need to look at;

http://docs.sun.com/app/docs/doc/816-5212/6mbcdgk3p?a=view

and tcp_ip_abort_cinterval

I'm not sure about another other versions or OSs.

Of course, I might be wrong...  :-)


On Mon, Mar 8, 2010 at 11:17 AM, Sim IJskes - QCG <si...@qcg.nl> wrote:

> Hello,
>
> does anybody know, or know how to calculate how long it would take (taking
> the complete system into account) from a call to a Remote to an error in the
> following scenario:
>
> A workstation TCP based Endpoint. Service registered. Remote handle
> retrieved by client. Physical network disconnect on server side.
>
> (From the TCP side off things this would constitute complete packetloss
> without other indications (no ICMP for instance, no interface takedown at
> client because of cable disconnect)).
>
> Gr. Sim
>
>
> --
> QCG, Software voor het MKB, 071-5890970, http://www.qcg.nl
> Quality Consultancy Group b.v., Leiderdorp, Kvk Leiden: 28088397
>