Re: Oracle and Raid setup

DA Morgan wrote:

>

No real statements in that article on differences in handling reads and writes in RAID 1, Daniel. In fact, due to the mix transaction workload and complete lack of detailed instrumentation in most (if not all) Unix environments, I'd say it's impossible to extract any info from the benchmarks - in what regards RAID intricacies, of course!

The Oak Table Network people have pointed out this problem a number of times, particularly in the latest books: there is a complete lack of detailed IO instrumentation to REALLY find out what is going on in them controllers/disks/fibre channels. In ALL OS environments (bar mainframes) and with just about all databases.

Other than benchmarking very specific loads, under very tight control, there is no way to find out exactly if a particular architecture is doing intelligent disk IO or not. And as soon as you restrict your benchmark load to just reads or just writes, you are not benchmarking a real world workload and all bets are out when it comes to bottom line: database performance!

Until this situation changes (don't hold your breath!), all we can really do is: measure indirect effects, read the source code when available, extrapolate from pure maths, experiment to find out if a theory matches a real result and try and adopt a solution that is fast and reliable.

Give me a proper mainframe IO subsystem anyday... Received on Sun Jun 12 2005 - 20:41:25 CDT