Re: SAME config

"David E. Grove" <david_grove_at_correct.state.ak.us> wrote:
> Thank you for confirming my impression.
>
> Can you also clarify the meaning of the 1 MB stripe width referred to in
> the SAME papaer? Does that have the frequent meaning of 1 MB per disk,
> or does it mean the sum of all the disks should be 1 MB (so that, in my
> case, each disk would be 1/6 MB [if using 6 disks]).

I am fairly certain that they mean 1 MB per disk.

> The former seems to be more consistent with the intent of the paper, but
> then what good is of making the transfer large enough so that the
> transfer time dominates, if I only need data from a single block on the
> disk?

Well, let's say you use 8K db-blocks, and 2K (versus 1M) stripe width. Let us further say the transfer time is 0.1ms/KB (just because that is what the SAME paper seemed to use) and that the positioning time is a random number uniformly from 0 to 20 ms, with an average of 10ms.

With the 1M stripe width, it takes on average 10 ms positioning time and 0.8 ms transfer time to read a 8K block, for 10.8 ms.

With the 2K stripe width (assuming you only need to read 4 spindles) you need to wait for the *worst* of the 4 independent random positioning times. That works out to 16 ms on average. But you save 0.6 ms tranfer by acting in parallel, so the total is 16.2 ms.

But this is assuming that there is no concurrent access. Now lets say there are 2 independent requests for two different blocks. If you use a large stripe with 5 spindles, there is only a 1/5 chance that both reads are trying to access the same spindle, resulting in head placement contention. If each block is spread over 4 spindles, there is nearly 100% chance that there will be head-placement contention. That will make the positioning time even worse.

Xho

-- 
-------------------- http://NewsReader.Com/ --------------------
Usenet Newsgroup Service                        $9.95/Month 30GB