V3 of PITR performance improvement for 8.4 (WIP)
This is the V3 of PITR performance improvement (readahead). The
change of the code is as follows:
1) Now readahead is integrated into the core so that it can deal with
sync.rep's log shipping.
2) posix_fadvise() call was integrated with Greg Stark's patch.
Still need some more work for the following:
1) Integrate with sync.rep's wal-receiver part,
2) fix to deal with WAL segment change.
V3 is posted for pre-review purpose and V4 will be posted at the
beginning of January.
Regards;
--
------
Koichi Suzuki
Attachments:
readahead-20081226.patchtext/x-diff; name=readahead-20081226.patchDownload+1276-87
"Koichi Suzuki" <koichi.szk@gmail.com> writes:
This is the V3 of PITR performance improvement (readahead). The
change of the code is as follows:1) Now readahead is integrated into the core so that it can deal with
sync.rep's log shipping.
2) posix_fadvise() call was integrated with Greg Stark's patch.
Wow, this is really cool. It integrates into core the readahead of WAL records
using a new RM method for each WAL type. That's great.
I haven't looked closely yet, I assume this avoids the code duplication? I did
notice there's a whole queue data structure for planning which blocks to
prefetch. Why is that necessary instead of just keeping a count of how many
blocks have been prefetched? Does it help avoid prefetching the same blocks
repeatedly?
--
Gregory Stark
EnterpriseDB http://www.enterprisedb.com
Ask me about EnterpriseDB's On-Demand Production Tuning