Re: Performance difference between 2 machines

"Sybrand Bakker" <postbus_at_sybrandb.demon.nl> wrote in message news:6hfeb11p8gu4m8jm9rtqpr7nprr8oq5ggo_at_4ax.com...
> On Mon, 20 Jun 2005 23:33:07 +0200, "Matthias Hoys"
> <idmwarpzone_NOSPAM__at_yahoo.com> wrote:
>
>>
>><sybrandb_at_yahoo.com> wrote in message
>>news:1119255391.716548.104290_at_g49g2000cwa.googlegroups.com...
>>> The obvious cause is the difference in database block size.
>>> Ct shouldn't have set it to 16k. Means one physical I/O executed by
>>> Oracle, will result in *2* physical I/Os on O/S level. Those two I/Os
>>> will suffer from rotational delay by the disks (ie, the second 16K will
>>> be read not immediately after the first).
>>> IMO, setting the database block size bigger than the O/S block size
>>> will result in massive delays... which is what we are seeing here.
>>>
>>>
>>> Hth
>>> ---
>>> Sybrand Bakker
>>> Senior Oracle DBA
>>>
>>
>>Sybrand, related question : we are using AIX 5.2 64-bit with jfs2
>>filesystems, data on EMC Symmetrix. What would be the recommended database
>>block size ? And do we need a smaller block size for OLTP than for DWH
>>databases ?
>>
>>Matthias
>>
>
> Howard Rogers on www.dizwell.com recommends 4k for AIX. I have no
> reason to doubt that, as that is straight from the AIX specific doco,
> and it is the AIX default.
>
> As to your second question: theoretically you would. However, remember
> it is of no use to set the block size smaller than a disk block. Which
> basically means 2k is ancient history. Remember also, you can speed up
> full table scans by setting db_file_multiblock_read_count. And setting
> the database block size affects both scattered reads and sequential
> reads, and Oracle will always read a complete block. This can hurt you
> in index lookups. If I recall Howard Rogers remarks, he is basically
> recommending to leave it alone, unless you have something Mickeysoft
> based. In that case any multiple of 8k will do.
>
>
> --
> Sybrand Bakker, Senior Oracle DBA

But, if I use a (low) block size of 4k, won't I risk ending up with lots of chained rows for "wide" tables (and thus double i/o to get those rows) ? Received on Mon Jun 20 2005 - 17:22:00 CDT