Re: Redo copy latch and redo allocation latch

You can also allocate multiple ( 2xCPU's set by log_simultaneous_copies ) redo copy latches on multi-processor machines. Then reduce redo_small_entry_max_size to a point where you begin to see more use of the copy latches. This will increase overall redo performance because you'll get more concurrency than when your threads ( user-sessions ) try to make use of the single redo allocation latch.

If you're running 7.3+, the redo_small_entry_max_size has been reduced by default on multi-processor systems to take advantage of this concurrency. I've set my small_size to 40 ( half that of the default of 80 ) and my simul.._copies to 16 on an 8xCPU UNIX platform and seen great improvement during heavy concurrent redo activity.

-Kevin

--

remove the x for an email reply
Jonathan Lewis wrote in message
<918637153.22348.1.nnrp-06.9e984b29_at_news.demon.co.uk>...
>
>David,
>A minor detail on your explanation -
>
>In a multiple CPU system the redo allocation latch
>is used by all processes to get their ration of the
>redo buffer very quickly; the redo allocation latch
>is then released and the redo copy latch is taken
>to do the actual copying into the allocated area.
>
>The copy latches are necessary to ensure that
>when a log buffer write takes place, LGWR does
>not write out a section of the log that has been
>allocated but not yet filled LGWR has to be able
>to grab all the redo copy latches before it is allowed
>to write.
>
>In answer to your question below - (and this is
>purely guesswork; I'll make a completely different
>guess if you decide that you described it the wrong
>way round) -
>
>On a single CPU system, there is no point in waiting
>for an allocation latch - if someone else has got it it
>is not going to come free whilst you spin, so you might
>as well give up immediately and release the CPU - this is
>not true, of course, on a multiple CPU system but maybe
>the code layer has not been modified.
>
>If you are using copy latches at all, you are on a multiple
>CPU system, and there is a reasonable chance that
>a latch will become free if wait a bit - it is cheaper therefore
>to waste a bit of CPU spinning rather than releasing your CPU
>and finding that you migrate to a different CPU when you
>wake-up again.
>
>Jonathan Lewis
>Yet another Oracle-related web site: www.jlcomp.demon.co.uk
>
>
>David Sisk wrote in message ...
>>
>>I have a box (NT) with 4 - 400Mhz Pentium Xeon processes (lots of extra
>>juice!), so I'm trying to shift more log buffer locking to use the redo
>copy
>>latches (by lowering the LOG_SMALL_ENTRY_MAX_SIZE parm). I'm seeing that
>>almost all Willing-to-wait latches are using a copy latch, while all
>>immediate latches are using an allocation latch. (Hmmmm, or do I have
this
>>backwards?)
>>
>>Any thoughts on this one?
>
>
>
Received on Wed Feb 10 1999 - 12:38:00 CST