Re: latches: cache buffers chains

I agree with JL, and think that at 99.7% you have other cats to fry, but....

I can add when vexed by buffer problems, I usually make hash buckets darn near 100% of block buffers.......and that in your case, you can kill sleeps ( which are expensive!) by upping the spin_count from the default ( 200 or so? ).
I have mine currently set to about 1000, and if i do miss a latch ( 10% of the
time! ), I virtually *always* get it _without_ sleeping.....

N.B.spinning does cost CPU cycles, but you have a better chance of picking up the latch in a timely fashion than if you sleep for it. Don't tune for CPU
usage, tune for response time and throughput.....

hth

Jonathan Lewis wrote in message
<926452390.1249.1.nnrp-02.9e984b29_at_news.demon.co.uk>...
>
>You have a good hit ratio, so I personally wouldn't worry
>too much about it.
>
>One thing to look at is the v$latch_children table for
>specific latches which individually responsible for
>misses (see my posts to Ross Mohan on 'an oldie ...'),
>then track the most expensive latch back to the
>database block it is protecting.
>
>You can be lucky and change the contention
>dramatically by changin the number of db block
>hash buckets so that very commonly used blocks
>(typically index root blocks) hash to different chains.
>
>Watch out - there are typically db_block_buffers/4
>child latches / hash chains.
>
>What is the spin_count ? You can get an idea on how
>much of your CPU is simply 'lost' in spinning by doubling
>and halving it, and seeing if this makes any significant
>difference to the CPU usage and total throughput.
>
>--
>
>Jonathan Lewis
>Yet another Oracle-related web site: www.jlcomp.demon.co.uk
>
>chuckh_at_safeplace.net wrote in message <7h9luu$85s$1_at_nnrp1.deja.com>...
>>I recently increased the buffer cache size to 300m on a 4 cpu NT Oracle
>>database. Since then I've noticed a lot more misses on the "cache
>>buffers chains" latch and a sleeps/miss ratio of 73.2%. What can I do
>>to reduce the number of misses on the latch and reduce the number of
>>sleeps? There seems to be no documentation on tuning this latch.
>>
>>The hit ratio on it is .997. Should I even be concerned about it?
>>
>
>
>
Received on Tue May 11 1999 - 22:24:22 CDT