Re: Multiprocessing: Oracle 7

From: James A. Lane <jl3a+_at_andrew.cmu.edu>
Date: Wed, 17 Aug 1994 22:37:36 -0400
Message-ID: <siIgZkG00WB8IJU7dG_at_andrew.cmu.edu>


>Up to the Informix standards ?? Oracle7 Release 7.1 is shipping on
>more then 30 or so platforms, which includes both SMP, Clustered and
>MPP systems. Informix 7.0 ships only on Sequent which is SMP, no clustering
>or MPP available until the end of '95 so I've heard.
>How can we beat that ??

You're confusing platform availablity with a complete architecture. Oracle can only run certain types of queries, generally run off a FTS, in parallel. Informix can run all in parallel. Oracle does things by process, not by thread: more overhead. The Informix release carries the parallel architecture all the way to the kernel, Oracle does not. This doesn't mean Oracle can't catch up, just that the product is presently on the back side of the technology leapfrog match. Also, Informix was available 6 months ago but limited due to a licensing agreement with Sequent.

Clustering in Informix I'm not sure about.

>I do not agree. The advantage of MPP systems is that the CPU power, but also
>memory bandwisth and I/O bandwith is expandable. If you need more I/O, just
>add I/O channels *AND* cpu's to handle this. So there's virtually no limit.

Assuming linear processor and I/O scalability. Not a good bet on MPP. And the MPP architecture is generally a shared-nothing architecture, meaning each CPU has its own address space, unlike a shared-memory architecture in an SMP server. The overhead involved is not small. There are latency problems with I/O going from a node to disk and back on MPP boxes. The solution (for Oracle) is the Gigacache, which takes advantage of the individual CPU memory to partition the MPP machine and basically run the database in memory.

Even with this, Oracle's benchmark on a Sequent vs. an nCube showed the problems with the MPP platform. A 20 proc Sequent beat the nCube with on the order of 160 procs. Even in FTS, where the nCube should be blasting away due to the parallelism and the fact that the database was practically all in memory. Don't confuse strong general vectorized CPU power with a well-designed commercial database server. Even assuming equivalent performance in the above configurations, you would be paying three times as much, if not more, for the same level of performance on the MPP box. Received on Thu Aug 18 1994 - 04:37:36 CEST

Original text of this message