Re: DB_MULTIBLOCK_READ_COUNT and DB_BLOCK_SIZE

Hello Venkat,

Look in v$system_event.
Every 'db file scattered read' is a multiblock read. Every 'db file sequential read' is a single block read. Let mc and sc be the respective counts, and let mt and st be the respective average times. Also let c be the multiblock read count. If you double c you would expect mc to be reduced and you would expect mt to be increased
because of the additional transfer time.

You can estimate the difference as follows: transfer time per block, tt = (mt - st)/(c - 1) new multiblock read time, mt' = mt + tt * c time saved, X = mt * mc - mt' * mc/1.9
(mc is reduced by less than a factor of 2 because of alignment issues).
As a percentage of time waiting for reads this is X/(mc*mt + sc*st).

Note, this estimate assumes that the increase in c is such that the multiblock read size is still less than the platform's max io size. On many platforms and versions, Oracle will not exceed the max io size, unless the use_readv parameter is set to true. Even then, the benefits of an increased multiblock read size is not so significant, but still worthwhile.

Hope this helps,
Steve Adams

On Sat, 19 Jul 1997 12:36:15 -0400, venkata terala <venkata_at_inetnow.net> wrote:

>Hi! Oracle gurus
>
>I am running 7.3.3.0.2 on Solaris 2.51. I need to know how much of a
>performance gain can one expect by increasing the DB_BLOCK_SIZE to 4 or
>8K and the DB_MULTIBLOCK_READ_COUNT to 16 or 32 for a highly read
>intensive and low-write intensive setup. I understand that the results
>vary with every situation - but can I expect atleast a low to moderate
>performance gain ? Has anyone seen a case where doing so actually
>degraded performance ?
>Or is my question too generic ?
>Any input will be greatly appreciated
>
>
>Venkat
>vreddy_at_wsg1.usa.com
Received on Mon Jul 21 1997 - 00:00:00 CDT