When I read the post related to the battery-backed disks, I come to this
question: What can we improve/reduce if we have them? Since our fsync
methods is already be able support that (thought not automatically), we may
want to look into other parts. One issue is the xlog and our aim is (1)
reduce xlog write requests/size (2) make commit faster.
The key is the number of pages changed by an atomic operation. For
operations change only 1 page, then we can assert that if you have written
the page, you are safe. So for buffer flush, if we know that this page is
the only page changed in atomic operations (no matter how many), we just
issue the write without xlog then it is done. For commit, we may have an
alternative to just write this page without any xlog flush. Of course, this
is a tradeoff, which may impose even bigger cost on commit due to
sizeof(page) vs. sizeof(xlog record), but we may win back the xlog replay
time.
Comments?
Regards,
Qingqing