[lvs-users] ipvs-dr and ip_vs_conn
horms at verge.net.au
Wed Nov 4 05:42:01 GMT 2009
On Tue, Nov 03, 2009 at 11:54:07AM +0300, calculator at land.ru wrote:
> Graeme Fowler:
> > Hi there
> > On Sun, 2009-11-01 at 11:10 +0300, calculator at land.ru wrote:
> >> Similar problem post on
> >> http://lists.graemef.net/pipermail/lvs-users/2005-May/013820.html
> > Do you mean you have a director which is running out of memory?
> No, my mistake.
> Records from /proc/net/ip_vs_conn are always removed and it's ok,
> but with default settings, sometimes i have "IPVS: ip_vs_conn_new: no memory available." with 4G memory...
> I don't understand: why if i use ipvs-dr scheme with sh
> sheduler there are records in /proc/net/ip_vs_conn?
> Does this is so because we do not want to drop "not full openen"/"not
> full closed" connection so we renew sh sheduler _routing table_?
I guess that technically it ought to be possible to do without
/proc/net/ip_vs_conn entries when the sh-scheduler is in use.
But from an implementation point of view this is quite
a departure from the way LVS works internally. That is,
the scheduler is called to choose a real-server for new connections
and an entry is created in /proc/net/ip_vs_conn which is used
for the rest of the life of the connection.
> And one more question. If I set 'ipvsadm --set 1 1 1' will it influence on
> whole ipvs (are packets which does not make whole cicle in 1 second
> droped?), or will it influence only on count of records in ip_vs_conn?
I doubt that timeouts that short will have any affect on the system
as there are other, per-connection-state timeouts which are significantly
larger. I imagine that the smallest workable timeout for TCP would
be around 2 minutes.
Are you sure that you are exhausting ~4bytes of memory
through the entries in /proc/net/ip_vs_conn? To put this
in perspective, this implies that you have in the
order of 30 million connections in /proc/net/ip_vs_conn.
More information about the lvs-users