Re: XLogInsert scaling, revisited

On Mon, Jul 8, 2013 at 1:38 PM, Heikki Linnakangas
<hlinnakangas(at)vmware(dot)com> wrote:
> GetXLogBuffer() doesn't read the content of the page - it writes to it (or
> rather, the caller of GetXLogBarrier() does). The barrier is needed between
> reading xlblocks (to check that the buffer contains the right page), and
> writing the WAL data to the buffer. README.barrier says that you need a full
> memory barrier to separate a load and a store.

Indeed you are right. My bad. I somehow thought that every location in
the WAL buffer is written once while it's actually done twice due to
AdvanceXLInsertBuffer() zeroing the page out.

Now thinking about it, if that memset or the full memory barrier in
GetXLogBuffer() ever prove to be a significant overhead, the
initialization could be optimized to avoid zeroing the page.
AdvanceXLInsertBuffer() would only write the header fields,
CopyXLogRecordToWAL() needs to take care to write out zeroes for
alignment padding and xlog switches and always write out xlp_rem_len
and XLP_FIRST_IS_CONTRECORD bit of xlp_info. xlp_info needs to be
split into two 8bit variables so XLP_FIRST_IS_CONTRECORD and
XLP_LONG_HEADER/XLP_BKP_REMOVABLE can be written into disjoint memory
locations. This way all memory locations in xlog buffer page are
stored exactly once and there is no data race between writes making it
possible to omit the barrier from GetXLogBuffer.
WaitXLogInsertionsToFinish() takes care of the memory barrier for
XlogWrite().

Regards,
Ants Aasma
--
Cybertec Schönig & Schönig GmbH
Gröhrmühlgasse 26
A-2700 Wiener Neustadt
Web: http://www.postgresql-support.de