On Thu, Nov 16, 2000 at 03:36:54PM +0000, Stephen Rowles wrote:
> I am running NAT LVS. This is used for a compute cluster and so has
> permanent connections almost 24/7.
>
> Currently there is not redundancy / failover provision for the director, so
> if it dies (for whatever reason) the cluster is out of action. I know that
> heartbeat can be used to enable another director if the main director
> dies... however, what will happen to the current connections? I assume that
> they will be lost as the failover director will not have a list of the
> current connections to manage.
This is correct.
> Is there anyway that the failover director can keep track of the current
> connections so that when it takes over the IP address, the LVS NAT
> forwarding can be taken over as well?
>
> I know it's a long shot but I thought it was worth a try.
Essentially what is needed is for information about the affinity for a
connection for a real server to be communicated to the standby. The trick
is that this has to be done without impacting on performance. Because if
this the information would most likely need to be communicated to the
standby in a non-blocking fashion. This is likely to create some
interesting race conditions. What if the active Linux Director crashes just
as a connection is established but the communitcation of this to the stanby
is still pending (possibly failed and is in the process of a retransmit).
Just some food for thought.
--
Horms
> I am running NAT LVS. This is used for a compute cluster and so has
> permanent connections almost 24/7.
>
> Currently there is not redundancy / failover provision for the director, so
> if it dies (for whatever reason) the cluster is out of action. I know that
> heartbeat can be used to enable another director if the main director
> dies... however, what will happen to the current connections? I assume that
> they will be lost as the failover director will not have a list of the
> current connections to manage.
This is correct.
> Is there anyway that the failover director can keep track of the current
> connections so that when it takes over the IP address, the LVS NAT
> forwarding can be taken over as well?
>
> I know it's a long shot but I thought it was worth a try.
Essentially what is needed is for information about the affinity for a
connection for a real server to be communicated to the standby. The trick
is that this has to be done without impacting on performance. Because if
this the information would most likely need to be communicated to the
standby in a non-blocking fashion. This is likely to create some
interesting race conditions. What if the active Linux Director crashes just
as a connection is established but the communitcation of this to the stanby
is still pending (possibly failed and is in the process of a retransmit).
Just some food for thought.
--
Horms