Re: Question on Parallel ORACLE on NCR 3600

From: John Susag <John.Susag_at_MinneapolisMN.NCR.COM>
Date: Tue, 26 Jul 1994 13:51:43 GMT
Message-ID: <CtJvu8.9C1_at_ncrcae.ColumbiaSC.NCR.COM>


>In article <311dre$t65_at_crl.crl.com> Carolyn Binder writes:
>Parallel Oracle Scenario (loosely coupled):
>Assume there are two processors on an SMP, sharing an Oracle database.
>If one requests a block of information that the other one has in memory,
>how will the data be obtained? I figured that the requesting process
>would end up getting the info from the memory of the other, rather than
>doing disk i/o, which'd be slower.
>
>The real reason I'm asking this is because I'm trying to understand how
>parallel Oracle works on an NCR 3600 box we're looking at. We want to
>reduce intraprocessor communication between APs because the Ynet is slower
>than an AP's internal bus. The reps are telling me that very little
>communication goes across the Ynet, and *no* database data travels on the
>Ynet. So in the case above, how do the processors exchange data? Far as
>I know, the YNET is what connects the two Aps; I think I'm missing something.
>
>Reply to either me or the newsgroup. Personally, I'd rather see more
>answers posted to the newsgroup, rather than emails to the posters; it
>makes for more interesting stuff to read! Thanks.
>
>Carolyn Binder cbinder_at_crl.com
>
>>
Carolyn, Oracle Parallel Server (OPS) is an example of "Scale-Up" parallelism. Multiple processors access the same data store, each performing it's own task. This allows you to use many more processors against the same data to scale up the number of transactions.

Since all these processors (across multiple nodes, linked by the 3600's Y-Net) are talking to the same data, you need a distributed lock manager (DLM) to maintain lock consistency. A process on one node requests a lock and the DLM uses the Y-Net to inform the other nodes of the lock request.

To my knowledge, OPS does not access cache across nodes. The only internode traffic will be lock requests which are very small.

Check out the TPC-A full disclosures on OPS. Most vendors link nodes with ethernet which is MUCH slower than the Y-Net.

John.Susag_at_MinneapolisMN.NCR.COM
AT&T Global Information Solutions
Opinions expressed are mine alone - Don't flame anyone else Received on Tue Jul 26 1994 - 15:51:43 CEST

Original text of this message