Re: lies damn lies and benchmarks

"Mark Townsend" <markbtownsend_at_attbi.com> wrote in message news:B9009833.20F3B%markbtownsend_at_attbi.com...
> in article YOzC8.717$eR3.46772_at_news.uswest.net, Pablo Sanchez at
> pablo_at_dev.null wrote on 5/9/02 12:08 PM:
>
> > The fact remains, that the unbiased TPC-C benchmarks show that
> > multiversioning and locking are non-issues.
>
> How do you arrive at this conclusion ?

Because if they were issues, the problems would have shown up early on in the 4-64p benchmarks. Recall when we first broke through the 100K tpmC's with V3 (right??) of the TPC-C. We were breaking that mark with 64p machines with RDBMS that were not using multi-versioning (Sybase ASE on HP).

> The TPC's clearly show that one vendor scales well to more CPU's in
a single
> box than any other vendor, and that the performance per CPU is also
stronger
> (not that that is a legal measure).

Actually, what Microsoft has done is to hit the 'sweet spot' in the database market: 4-8P machines. The percentage of customers in the above 32p range is very few and they don't care. It's the way it ought to be.

> This is why clustering has destroyed the TPC-C as a valid measure
(in much
> the same way as MV's destroyed the TPC-D).
>
> Simply put, any vendor that can't scale up can now scale out and the
vendor
> that can get the biggest kit together wins. Look at the disk
alone - one
> result uses 60 terabytes of disk per node, for 32 nodes. Over 3500
18 Gig
> disk drives - without mirroring. What's the average MTTF for an 18
Gig
> drive - once a year - twice a year ?

About five years.

> In this config, 5-10 drives would be
> failing daily. Recovery would be a nightmare, and with 128 Terabytes
of
> memory on board, it would take days to re-load all the caches - but
> nevermind that nobody in the real world can take advantage of all
this
> effort.

I believe this is a red herring since the benchmark isn't about recovery, it's about throughput and a competitive 'price/tpmC' If the benchmark was supposed to measure recoverability costs, then it'd be addressed as a metric.

> Time for a new benchmark. One that measures price/performance AND
> availability.

Interesting point. I like that new metric.

Are you the Mark Townsend of Oracle? Why not lobby your TPC-C representatives to make the motion at the -C meeting for the metric? I think it has a lot of merrit.

--
Pablo Sanchez, High-Performance Database Engineering
mailto:pablo_at_hpdbe.com
http://www.hpdbe.com
Available for short-term and long-term contracts