On Thu, Jul 5, 2012 at 12:13 PM, Reza Taheri
<rtaheri@vmware.com> wrote:
Hi Robert,
Yes, the same concept. Oracle's IOT feature is used often with TPC benchmarks.
Reza, it would be very helpful if you were to provide the list with a lot more information about your current software and hardware configuration before coming to the conclusion that the only possible way forward is with a significant architectural change to the db engine itself. Not only is it not at all clear that you are extracting maximum performance from your current hardware and software, but I doubt anyone is particularly interested in doing a bunch of development purely to game a benchmark. There has been significant discussion of the necessity and viability of the feature you are requesting in the past, so you should probably start where those discussions left off rather than starting the discussion all over again from the beginning. Of course, if vmware were to sponsor development of the feature in question, it probably wouldn't require nearly as much buy-in from the wider community.
Getting back to the current performance issues - I have little doubt that the MS SQL benchmark was set up and run by people who were intimately familiar with MS SQL performance tuning. You stated in your earlier email that your team doesn't have significant postgresql-specific experience, so it isn't necessarily surprising that your first attempt at tuning didn't get the results that you are looking for. You stated that you have 14 SSDs and 90 spinning drives, but you don't specify how they are combined and how the database is laid out on top of them. There is no mention of how much memory is available to the system. We don't know how you've configured postgresql's memory allocation or how your config weights the relative costs of index lookups, sequential scans, etc. The guidelines for this mailing list include instructions for what information should be provided when asking about performance improvements.
http://archives.postgresql.org/pgsql-performance/ Let's start by ascertaining how your benchmark results can be improved without engaging in a significant development effort on the db engine itself.
