> >>>>> "Mark" == Mark A <maswitchboard.net> writes:
> Mark> "Sailesh Krishnamurthy" <seesignaturecs.berkeley.edu> wrote
> Mark> in message news:bxy1xx3um5o.fsfdatafix.cs.berkeley.edu...
> >> That sounds about right to me - I don't know of any other
> >> commercial system that had shared nothing support before
> >> Teradata. Of course, Teradata's market is ytics as opposed
> >> to transaction processing.
> Mark> I don't think the first release of DB2 Parallel Edition (as
> Mark> it was called back then) was much different than Teradata in
> Mark> its focus on queries rather than transaction processing. I
> Oh, I guess it wasn't but now I'm not speaking for IBM - actually I
> rarely am speaking for IBM. Let me amend that - I never speak for IBM,
> and given that IBM never speaks for me that's fine :-)
> In fact, in general I'm curious as to people's experiences with shared
> nothing parallelism in OLTP environments. Certainly the DB2 on Windows
> TPC-C benchmarks show that it's possible to scale well. Still I'm
> curious to learn more about "war stories". Just out of academic
> interest !
> Mark, I'm not opposed to your response to Spencer that a
> shared-nothing cluster database was/is not designed for a large number
> of vanilla insert statements. That said, it's certainly interesting to
> get a handle on why we see some performance results.
> I can understand Spencer's concerns. However, what will now be
> interesting is if he can post a comparison of the insert times he can
> generate with buffered inserts in serial and parallel mode. Then we
> can better test the theory that what he's measuring is parallel