Oracle FAQ Your Portal to the Oracle Knowledge Grid

Home -> Community -> Usenet -> comp.databases.theory -> Re: Designing database tables for performance?

Re: Designing database tables for performance?

From: paul c <>
Date: Wed, 07 Mar 2007 00:06:15 GMT
Message-ID: <XlnHh.1239650$1T2.246124@pd7urf2no>

joel garry wrote:

> On Mar 6, 12:52 pm, "Cimode" <> wrote:

>>On 6 mar, 05:59, "" <> wrote:
>>>On 24 Feb, 13:30, "Cimode" <> wrote:
>>>>On 23 fév, 22:33, "jgar the jorrible" <> wrote:
>>>>>>In what RAM would be less physical than HD ? For any reason, an
>>>>>>absurdity is an absurdity.
>>>>>Not an absurdity, you just aren't paying attention to how the I/O is
>>>>So you say there are *ways* to count IO's. Fair enough. Question is:
>>>>what has the way of counting IO's has any bearing on the media that
>>>>supports them and therefore qualifies their nature as physical or
>>>>logical? What is the difference: speed?
>>> Yes. A logical I/O does not go down an I/O channel but comes via the
>>>memory bus.
>>And in what a memory bus is less *physical* than an IO channel? How
>>do you think such bus is filled with data at some point in time
>>otherwise than by a pull on the IO channel.
> ??? The whole point is that it might have been updated only in memory
> and may have nothing to do with I/O until some time in the future, or
> maybe never if it is rolled back.
> ...

No, that's only one aspect of the point. When Codd spoke of assymetry, he certainly had performance in mind. Then he advocated a limited version of relations that he could see could compete with the hardware-oriented approaches of the day.

All IO's aren't equal, eg., process-to-process transfers might be an order of magnitude cheaper than disk I/O. How much depends on the algorithms and the concepts the engine tries to implement. RT is not immune to this.

Today, many, perhaps a majority of programmer operate under the illusion that memory "IO" is of no time cost. In fact, it often reduces the CPU speed people think they bought by an order of magnitude.

On the other hand, at one time in the 1970's, every 28 days, I had to consolidate the chocolate bar sales of the provinces of Canuckistan. This required much disk IO and sorting and merging. That was with a machine that had only 64K of main memory. I used to dream of 256K - with that I could have done the whole so-called country in one go, saving much setup time as well as other chances for errors.

Later, I had access to mainframes that had as much as 16MB "real" (ha, ha) memory and was completely pissed off that the OS designers stole much of that for their own buffers of various kinds (some were even buffers for program code).

Later still, when memory approached 1MB, the same damn thievery continued. But I was still telling programmers using a two-process architecture to count their logical accesses as one cent, their process transfers as ten cents and their physical IOs as a buck. (The ones that refused to do this got crappy reviews.)

Now, memory latency is actually a bigger problem, relatively, than disk latency ever was, but the Wintel coalition doesn't talk about it much. Meanwhile, the essential aspects of the census for small monarchies like Canuckistan could be handled in "main memory" of a consumer machine.

>> ...
>>So because time is the difference that makes less physical. Don't you
>>see anything wrong in that?

> ...

Don't see anything wrong, but as far as db is concerned, Cimode's observation seems accurate to me.

p Received on Tue Mar 06 2007 - 18:06:15 CST

Original text of this message