Re: Larry Ellison comments on Microsoft's benchmark

"Blair Kenneth Adamache" <adamache_at_ca.ibm.com> wrote:
>
> Ivana, has OPS been used to publish a recent TPC-C benchmark? The top
> ten only seems to have Oracle 8i and Oracle 7 results. The only OPS
> result I can find is from 12/98, somewhere around 37th place (shown
> below).

"OPS" is not a separate product from "Oracle 8i". Rather, OPS is an option of Oracle8i. So unless I am mistaken, whenever Oracle 8i is used in a cluster, the Oracle Parallel Server is being used.

> If could beg a small indulgence, Ms. Humpalot, does OPS run into
> some scaling difficulties because of shared disk? My view of HA
> on shared nothing is that if you have 12 nodes, each node can be
> backed up in a mutual takeover configuration by 1-11 other nodes
> (depending on whether you cluster for HA pair wise, or use
> something more complex).

The above appears to conflict with what you said earlier. (Following paragraph reproduced from your earlier post:)

> The published benchmark does not include high availability machines or
> Microsoft Cluster Server (MSCS). If you wanted to back up each machine,
> you'd need hot standby machines (doubling the number of CPU's and RAM,
> but not increasing the disk), mutual takeover machines running some other
> workload, or have the machines back each other up (which is very common
> for DB2 EEE customers).

So first you said you have to double the number of CPU's and RAM. Now you appear to be advocating a different approach -- rather than doubling the CPU's and RAM, each node can be backed up in a mutual takeover configuration?

Note that a mutual takeover situation still has at least 6 times the failure rate as a real cluster in which if one machine fails the other machines continue to work uninterrupted, balancing the load.

Still, I am interested in your theory that in a shared-nothing database, if one machine fails other machines can continue to access the failed machine's disk. Could you provide additional technical details on how this is possible? My understanding was that the disk and the data on it will become inaccessible.

> No need for a distributed lock manager to resolve who owns the
> disk. NODE A dies - NODE B takes over. In shared disk, NODE A
> dies, and NODE B fights with other nodes using that disk to save
> the work that NODE A was doing.

In Oracle Parallel Server, all machines mount the entire database. There is no issue of one machine taking a portion of the database with it when it crashes. The disks are a central resource, seen as a single entity by the machines in the cluster. If one machine crashes there is no impact on the disks. Received on Thu Jul 06 2000 - 00:00:00 CDT