Heikki Linnakangas wrote:
>> Why does WAL replay read much more than it writes?
>> I thought that pretty much every block read during WAL
>> replay would also get dirtied and hence written out.
>
> Not necessarily. If a block is modified and written out of the buffer
> cache before next checkpoint, the latest version of the block is
already
> on disk. On replay, the redo routine reads the block, sees that the
> change was applied, and does nothing.
True. Could that account for 1000 times more reads than writes?
>> I wonder why the performance is good in the first few seconds.
>> Why should exactly the pages that I need in the beginning
>> happen to be in cache?
>
> This is probably because of full_page_writes=on. When replay has a
full
> page image of a block, it doesn't need to read the old contents from
> disk. It can just blindly write the image to disk. Writing a block to
> disk also puts that block in the OS cache, so this also efficiently
> warms the cache from the WAL. Hence in the beginning of replay, you
just
> write a lot of full page images to the OS cache, which is fast, and
you
> only start reading from disk after you've filled up the OS cache. If
> this theory is true, you should see a pattern in the I/O stats, where
in
> the first seconds there is no I/O, but the CPU is 100% busy while it
> reads from WAL and writes out the pages to the OS cache. After the OS
> cache fills up with the dirty pages (up to dirty_ratio, on Linux), you
> will start to see a lot of writes. As the replay progresses, you will
> see more and more reads, as you start to get cache misses.
That makes sense to me.
Unfortunately I don't have statistics in the required resolution
to verify that.
Thanks for the explanations.
Yours,
Laurenz Albe
