Checkpoints perf improvment in 4.4 vs 3.2

Question

Hi,
Can someone tell me wether there has been any perf improvment done in the checkpoint processing between 4.4 and 3.2 releases.
If yes, I'm interested in getting some details/pointers onto what has been done.
If no, are there any plans to do some perf improvments ?
Regards,
Jerome

Answer 1

hello,
I checked into this further and some additional information
follows. If there is something in particular that you are
interested in, let me know.
In general there have been a variety of checkpoint improvements
over the years, mostly concentrating on keeping checkpoints from
impacting normal work processing and in avoiding writing pages
that don't need to be written to implement transactional guarantees.
The buffer manager was significantly rewritten in 4.1, which
included how checkpoints work. The specific change there was to have
the checkpoint code always complete. In previous releases the
checkpoint could return DB_INCOMPLETE if the checkpoint thread
could not get control of a buffer due to high traffic on that
buffer by other threads.
After 4.1 the buffer manager used multiple mutexes rather than a
single mutex, so access to the buffer pool by other threads should
be greater during a checkpoint than it was previously.
SR 11031 addressed a problem where a mutex was held in the buffer
manager while the fsync system call was issued locking out some types
of accesses to the buffer pool.
The next release will have additional changes for checkpointing. There
are improvements in the way dirty buffers are located during a
checkpoint which can have noticeable improvements with large buffer
pools with a majority of clean buffers.
thanks,
Sandra