RE: RAC - mix Intel and AMD in same cluster ?
- From: "Matthew Zito" <mzito@xxxxxxxxxxx>
- To: <Mark.Bobak@xxxxxxxxxxxx>, <Jon.Crisler@xxxxxxx>, "oracle-l" <oracle-l@xxxxxxxxxxxxx>
- Date: Thu, 23 Jul 2009 17:50:29 -0400
I certainly believe that it is possible, especially in highly
parallelized environments, for performance to get gated at the speed of
the slowest node. However, I definitely have customers in OLTP
environments where they added faster nodes (i.e. 4 nodes at 2 GHz, 2
nodes at 3.2Ghz), and saw the overall throughput improve, and saw the
two new nodes with a higher number of sessions, but comparable SQL
response time.
I don't know enough about the lock hashing to say for certain, but lock
management is definitely a high-priority event, so assuming your newer
nodes indeed take up a proportionately higher number of
sessions/queries, lock management should scale appropriately.
This would be an interesting benchmark factory experiment. I might have
some free time and free hardware late next week to give this a shot.
Thanks,
Matt
> -----Original Message-----
> From: oracle-l-bounce@xxxxxxxxxxxxx
[mailto:oracle-l-bounce@xxxxxxxxxxxxx]
> On Behalf Of Bobak, Mark
> Sent: Thursday, July 23, 2009 5:42 PM
> To: Jon.Crisler@xxxxxxx; oracle-l
> Subject: RE: RAC - mix Intel and AMD in same cluster ?
>
> Hi Jon,
>
> I had a similar concern recently, as we have a 3-node cluster of
DL-585
> G2s, and we were looking at adding two DL-585 G5s. The advice I got
was
> that it may not be the best idea. Certainly, it will work, that's not
the
> issue. And the question isn't really AMD vs. Intel, it's more about
the
> CPU capacity of the newer nodes vs. the older nodes, and how many
cores
> you're running, in terms of the number of CPU licenses. The issue is
> that, due to the way that RAC works, the performance of the newer
nodes
> will be gated at the performance of the older, slower node. So, in
our
> case, we've got 3 G2s w/ 4 dual core CPUs, and we were looking at
adding 2
> G5s with 4 quad-core CPUs. Due to the extra licensing cost of the two
> newer nodes, and the limited performance of the nodes when in a
cluster w/
> slower performing nodes, it's actually cheaper to get 3 new G5s and
re-
> purpose the G2s for something else.
>
> The reasoning behind why faster nodes in a cluster are gated to the
> performance of the slowest node, is that all nodes potentially
participate
> in any SQL execution, even if it's not parallel. That's because lock
> mastering of data blocks is hashed to all the nodes. So, as soon as
one
> of the slower nodes starts having CPU problems, lock management will
> suffer across the cluster, and everybody slows down. So, even though
you
> may have a couple of nodes capable of much more CPU load, their
> performance suffers due to the fact that they must participate in the
> cluster, and the cluster is limited by the slowest node.
>
> Hope that helps,
>
> -Mark
>
> -----Original Message-----
> From: oracle-l-bounce@xxxxxxxxxxxxx
[mailto:oracle-l-bounce@xxxxxxxxxxxxx]
> On Behalf Of Crisler, Jon
> Sent: Thursday, July 23, 2009 4:54 PM
> To: oracle-l
> Subject: RAC - mix Intel and AMD in same cluster ?
>
> Has anybody here built or upgraded a 10g RAC cluster using a mix of
both
> Intel and AMD processors ? We have an existing 3 node RAC cluster, RH
> Linux 4 with AMD Opteron 3 ghz dual core cpu's. It is being
considered
> to add a 4th node but with Intel CPU since the servers we used ( HP DL
> 585 G2 ) are no longer available new. Any thoughts in mixing Intel
and
> AMD processors in the same cluster ? We are also looking at upgrading
> the existing processors from dual-core to quad-core, but the quads
seem
> to max out at 2.4 ghz while the duals go to at least 3.2 ghz.
Obviously
> OS and Oracle will stay same version, but we might be forced to use
> Intel to add additional nodes, assuming it even works.
> --
> //www.freelists.org/webpage/oracle-l
>
>
>
>
> --
> //www.freelists.org/webpage/oracle-l
>
--
//www.freelists.org/webpage/oracle-l
Other related posts:
