Re: RCubes

From: Bjorn Bentzen <bjorn.bentzen_at_lindorff.com>
Date: 19 Dec 2002 01:25:46 -0800
Message-ID: <c5d27d1a.0212190125.154ac75b_at_posting.google.com>


"abombss" <abombss_at_hotmail.com> wrote in message news:<BgaM9.220948$pN3.17894_at_sccrnsc03>...
> > Bjorn Bentzen Wrote:
> > You can expect performance increases in the range 3-20 times for most
> > DSS-type of queries. IQ uses a "snapshot versioning" approach for
> > changes to data - an update of 10 million rows will be almost as fast
> > as an insert, with no special rollback/log-area to consider.
>
> How often do you run your ETL's, sounds like they run quick will little
> overhead?
>

On our current configuration, we run ETL's monthly.(IQ is only used for a portion of our data waehouse today) However, we are in the process of converting our existing warehouse to IQ, and the loads will become daily and monthly in the future.

ETL's for our monthly runs are not epsecially fast - this is not due to IQ, but rather the extensive transformations occuring in the ETL-stream (key-lookups for instance). Wherever we can, we use views in IQ instead of ETL-processing, because IQ is a lot faster.

> > There is no maintenance of indexes in IQ - you do not have to run
> > "update statistics" in order for the index to be useful. IQ allows
> > many types of indexes, depending on what the column is used for, e.g.
> > low-cardinality, medium-cardinality, summarised columns, join columns.
> > This means that you have to have a good grasp for what the data is
> > used for up front, but later on you do not need to maintain these
> > indexes in any way. IQ compresses data before it is written to
> > disk-with only the default index in place - the compression is 40-60%.
> > Even with specialised indexes on some columns, you can be sure that IQ
> > will only use a fraction of what a traditional RDBMS would use.
>
> This sounds great, I have been exploring solutions with Oracle 9i, but the
> maintence overhead seems huge. Dropping indexes for every ETL,
> recalculating statistics, partioning tables and indexes properly. And after
> all that stuff, one unanticipated ad-hoc query could still take hours to
> run.
>
> > IQ Multiplex is an architecure which allows IQ to scale almost 100%
> > linearly by putting in a new SMP_server box. This architecture is
> > dependent on a SAN-type of storage solution.
>
> What kind of hardware are you running?
> How big is your warehouse?
> What is pricing and licensing like for IQ?

Today we are running IQ and Sybase ASE on one 4-CPU NT-server, 4GB RAM. This will be expanded to 2 4-CPU Win2000 servers in a SAN, using raw devices. We expect an order of magnitude better perfomance from this.

The warehouse on IQ today is about 80GB, to become 300GB. Largest table is 70 million rows - 50 columns wide. One of the nice things about IQ is that is doesn't really matter how wide a table is, because of the column-orientation.

As regards pricing, I can only say that IQ is CPU-priced, not very expensive (compared to alternatives) and easily scalable by putting in a new server when needed (no hassle).

>
> Thanks for the outstanding feedback!
>
> Cheers,
> Adam

Regards,
Bjorn Bentzen Received on Thu Dec 19 2002 - 10:25:46 CET

Original text of this message