Re: [gentoo-cluster] gentoo 2006.1 + openmosix-sources-2.6.12.r577 + openmosix-user?

[Ramon van Alteren]

Hi,

momentics wrote:
> (cannot be agreed that grids needs to be explicitly defined, but we
> can leave it as is â?? that's not important actually)
Nope, true. I wasn't actually looking for a explicit definition of a
grid, the description below is just fine.
> Your example is quite competitiveâ?¦ I think that it works well when
> properly used.
> btw I'm working at dell emea :) - glad to hear of dellserv using ;)
Ah, well we're happy with the hardware, the only complaint is that it is
quite power-hungry ;-)
>
> Well, we are using 3dns/bigip for web/application/etc layers. I'm
> expert here so there are no questions. The question is all about of
> oracle rac that uses LVS in round robin (LVS is configured directly on
> database nodes).
>
> The actual question
>
> Our development and app management is divided on two parts: "the
> database" and "the rest".
> Oracle/linadm are using LVS when they are building databases for us.
> (just to clarify the question) Oracle RAC consists of several
> Active/Active nodes, each of them has its own node address. There is
> also dataguard but it is out of scope in this context.
>
Clear so far.
> So, we have, say, N nodes and N node IPs. Normally, these IPs I should
> use in Oracle client TNS name on app boxes. But linux gents are
> building LVS on top, where we additionally have K LVS' IPs (K=N) and
> they are suggesting to use LVS' instead of Oracle's ones. That's
> probably is fine.
> Oracle clients on each app box is configured in LB mode. Transactions
> are short lived then connection is recycled (in a test period I'm
> speaking of â?? immediately)

>
> So then, I'm using MRTG to see database loadâ?¦. Yes! It is not
> loadbalanced properly over time.
> When I'm injecting F5 that points to N database nodes in roundrobin
> and proposes 2 BigIP's IPs in the same OraClient configuration it is
> loadbalanced properly.
This is where it is getting strange.
You are using K LVS ip addressen and you say that K=N (amount of nodes) ???
That would mean that you have a loadbalancer address for every node ?

We are using two loadbalancers with a single ip for each database cluster.
The mysql-setup we're using is a simple replication setup meaning that
the loadbalanced nodes are read-only with a writeable master. The nodes
are reachable through LVS which we have setup using a Direct Routing
approach.

Previous experience with both our webfarm and database farm suggests
that the LVS rr algorithm is not the most efficient algorithm. We
switched to using wlc (weighted least connections) because of a number
of reasons.

The most important being:
Queries are not made equal, one query will occupy a database server for
far longer than the other.

We've found that the amount of open connections on a database server is
a far better indicator of it's load.

Apart from that you might want to take a look at the stickyness setting
(persistence) of the loadbalancer, this influences any loadbalancing
algorithm because it forces the loadbalancers to keep a certain client
with a previously allocated node if it returns for a query within a
time-threshold. With LVS if not specified the default is 300 seconds
which is quite long. Due to the nature of mysql-connections we use no
persistence at all on our database loadbalancers.

If you want to start using wlc or lc you should be aware that it is
dependent on proper fault detection on the nodes. Because it distributes
the load according to the amount of connections a broken server (with 0
connections) will suck up all new connections to the loadbalancer. We're
using keepalived to check on our realservers and remove them from the
loadbalancing pool in case of problems.
>
> The question is â?? I'm trying to find at least 1 idea on why we need to
> use LVS here.
>
I don't think there is a specific need, you could probably just as well
use the F5 loadbalancers. Depends on your budget probably ;-)

Regards,

Ramon

-- 
gentoo-cluster@g.o mailing list