On Wed, 2010-02-17 at 16:40 -0700, hj lee wrote:
>
> On Fri, Feb 12, 2010 at 11:17 PM, Steven Dake <[email protected]>
> wrote:
>
> On Fri, 2010-02-12 at 12:51 -0700, hj lee wrote:
> > Hi,
> >
> > If there are only two nodes in cluster and their IP
> addresses are
> > known a prior, then isn't it better to use TCP as a
> transport layer?
> > With heartbeat, there is a way to configure nodes before
> starting
> > cluster. Is TCP ever considered in corosync?
> >
>
>
> Not sure why tcp would be a better transport layer for two
> nodes.
> TCP/IP's key driving design factor from Darpa was to remain
> operational
> and _mask faults_ (not detect faults) under nuclear attack
> where many
> network links and routing systems would be under considerable
> changing
> stress. As a result TCP/IP is very resilient to faulty
> networks and
> packet loss but does not provide suitable fault detection.
> Further
> there is not automatic node discovery in TCP/IP. In short,
> TCP/IP while
> highly versatile doesn't offer the best characteristics for
> cluster
> communication.
>
> Finally, Corosync is designed for nway redundant cluster
> configurations.
> The 2N model is a simplification of the nway redundant model
> and we
> don't provide special behaviors during 2N operation.
>
> Regards
> -steve
>
> Actually I changed the code to use TCP just for sending a token, it is
> working very well. "very well" means I do not see token timeout any
> more. I know this is an ugly hack! The main reason I want to use TCP
> is I am seeing token lost timeout in heavy load, so the cluster is
> divided for very short time, which caused some problem in my
> application. If I run the system over night, I usually see one or two
> token lost timeout. the easy fix will be increase token timeout, but I
> have a strict requirement on timeout, so I couldn't increase it. So I
> tried to use TCP. After adding TCP transport just for token transmit,
> this timeout does not happen any more.
>
> The token is transmitted by unicast, so this chagne will work with
> more than two nodes. And I think there will be cases or environment
> this TCP token transmit may be useful or work better. At least it
> solves my case.
>
> Thanks
> hj
>
>
Did you try increasing (from man page):
token_retransmits_before_loss_const
This value identifies how many token retransmits
should be
attempted before forming a new configuration. If this
value is
set, retransmit and hold will be automatically calculated
from
retransmits_before_loss and token.
The default is 4 retransmissions.
_______________________________________________
Openais mailing list
[email protected]
https://lists.linux-foundation.org/mailman/listinfo/openais
