Re: Larry Ellison comments on Microsoft's benchmark

For what it's worth, the top TPC-C result is now DB2 on NT, another shared nothing database (unlike Microsoft SQL Server, this version of DB2 will allow the partition key to be updated). See: http://www.tpc.org/new_result/ttperf.idc. Between TPC-C and TPC-H (http://www.tpc.org/new_result/h-ttperf.idc) IBM software (DB2) and/or IBM hardware (a mixture of Netfinity, RS/6000 and NUMA-Q) now hold top spots for all TPC-C and TPC-H metrics that focus on performance.

Serge Rielau wrote:

> Finally this thread made to the DB2 newsgroup, eh?
>
> Here are my 2 (biased) cents:
> 1. Microsoft was sued over that benchmark because they violated one of
> the rules.
> I.e. SQL Server cannot update the column used to partition the view
> over the
> federated database. The TPC-C benchmark requires updateability of ALL
>
> columns. It seems like they'll get away with flagging their violation
> and a raised
> finger.
> To be fair I should add that updating of partitioning keys is no
> trivial excercise.
> 2. The benchmark did not use mirroring. As stated in earlier posts
> running such a
> beast in a company would be quite - unstable. One has to watch this
> when
> looking at the price/performance numbers.
> 3. Jim Gray said himself that the environment was very hard to set up
> and to keep
> running through the audit.
>
> Finally a federated database is not the same as an MPP system like e.g.
> DB2 EEE.
> In an MPP system the whole query plan gets compiled with MPP in mind and
> parts
> of the execution get distributed to the participating nodes. The whole
> thing is still one database, partitioned tables are still tables and the
> integration is VERY tight.
> A federated database sits on top of other database systems. Parts of the
> query get
> shipped (like SQL Servers pass through queries) to the target systems
> and the
> results get shipped back. On DB2 side this would be Datajoiner or the
> new DB2 V7.1 where SQL queries get reverse engineered post optimization
> and send to the target systems through public interfaces. The connection
> is loose compared to MPP and involves sending the SQL (rather than
> "executable sub query plans"). Partitioned tables are represented as
> views with all their advantages and disadvantages.
>
> just my two cents
> Serge
Received on Tue Jul 04 2000 - 00:00:00 CDT