Re: ORACLE ON VAX CLUSTERS
Date: 15 APR 93 13:24:56
Message-ID: <1993Apr15.174000.22488_at_ryn.mro4.dec.com>
In article <2767.1599.uupcb_at_factory.com>, jared.hecker_at_factory.com (Jared Hecker) writes...
>JF>How does this work. At what point do added nodes degrade the performance
>JF>of Oracle as it has to manage locks for many CPU's?
>
>James -
>
>Basically you should not be using VAXclusters for transactions processing
>systems as, last time I looked, Oracle was still using the CI bus to manage
>buffer cache update requests.
Oracle Parallel Server V6, the first version, did not use the CI bus to manage buffer cache update requests. When data from one cache needs to be shared with another node, it is flushed to the disk. The requesting node reads the data from the disk. Rdb/VMS uses essentially the same technique in a VMScluster.
Interestingly, the Parallel Server on the nCUBE platform does not bounce the data off the disk. Rather it uses the hypercube interconnect to transfer a buffer cache from one processor (node) to another.
I'd be interested to hear if things have changed for Oracle Parallel Server V7 in the VMScluster environment.
> As to where the point of diminishing marginal
>performance is, that is very much a function of application design and data
>being accessed. I don't think I'd want to run more than eight nodes doing
>this,as CI traffic can get pretty congested in this type of environment.
As usual, it all depends on the application. If the application can be well partitioned such that partions do not share much data, then a VMScluster is an excellent and extremely powerful platform for transaction processing. The less well the application can be partitioned, the lower the scaling as nodes are added to the cluster. A transaction processing monitor such as VAX ACMS can be used to route transactions to the appropriate database servers for processing.
In any case 8 nodes in a cluster is a lot of horsepower. A single DEC 1000 AXP system (a uni-processor _at_200MHz) has achieved 258 tpsA at a cost of k$ 8.5 per tpsA. This is the highest performance achieved to date on a uni-processor system.
In addition to performance, VMSclusters offer high availability and
automatic failover and recover features that are very valuable
to a TP shop.
>
Robert K. Abbott
abbott_at_tps.enet.dec.com Received on Thu Apr 15 1993 - 13:24:56 CEST