Speedup twophase transactions

From: Stas Kelvich <s(dot)kelvich(at)postgrespro(dot)ru>
To: pgsql-hackers <pgsql-hackers(at)postgresql(dot)org>
Subject: Speedup twophase transactions
Date: 2015-12-09 18:44:28
Message-ID: 74355FCF-AADC-4E51-850B-47AF59E0B215@postgrespro.ru
Views: Raw Message | Whole Thread | Download mbox | Resend email
Lists: pgsql-hackers


While working with cluster stuff (DTM, tsDTM) we noted that postgres 2pc transactions is approximately two times slower than an ordinary commit on workload with fast transactions — few single-row updates and COMMIT or PREPARE/COMMIT. Perf top showed that a lot of time is spent in kernel on fopen/fclose, so it worth a try to reduce file operations with 2pc tx.

Now 2PC in postgres does following:
* on prepare 2pc data (subxacts, commitrels, abortrels, invalmsgs) saved to xlog and to file, but file not is not fsynced
* on commit backend reads data from file
* if checkpoint occurs before commit, then files are fsynced during checkpoint
* if case of crash replay will move data from xlog to files

In this patch I’ve changed this procedures to following:
* on prepare backend writes data only to xlog and store pointer to the start of the xlog record
* if commit occurs before checkpoint then backend reads data from xlog by this pointer
* on checkpoint 2pc data copied to files and fsynced
* if commit happens after checkpoint then backend reads files
* in case of crash replay will move data from xlog to files (as it was before patch)

Most of that ideas was already mentioned in 2009 thread by Michael Paquier http://www.postgresql.org/message-id/c64c5f8b0908062031k3ff48428j824a9a46f28180ac@mail.gmail.com where he suggested to store 2pc data in shared memory.
At that time patch was declined because no significant speedup were observed. Now I see performance improvements by my patch at about 60%. Probably old benchmark overall tps was lower and it was harder to hit filesystem fopen/fclose limits.

Now results of benchmark are following (dual 6-core xeon server):

Current master without 2PC: ~42 ktps
Current master with 2PC: ~22 ktps
Current master with 2PC: ~36 ktps

Benchmark done with following script:

\set naccounts 100000 * :scale
\setrandom from_aid 1 :naccounts
\setrandom to_aid 1 :naccounts
\setrandom delta 1 100
\set scale :scale+1
UPDATE pgbench_accounts SET abalance = abalance - :delta WHERE aid = :from_aid;
UPDATE pgbench_accounts SET abalance = abalance + :delta WHERE aid = :to_aid;
PREPARE TRANSACTION ':client_id.:scale';
COMMIT PREPARED ':client_id.:scale';

Attachment Content-Type Size
2pc_xlog.diff application/octet-stream 12.8 KB
unknown_filename text/plain 95 bytes


Browse pgsql-hackers by date

  From Date Subject
Next Message Robert Haas 2015-12-09 19:13:13 Re: Redundant sentence within INSERT documentation page (exclusion constraints)
Previous Message Robert Haas 2015-12-09 18:35:35 Re: Unicode collations in FreeBSD 11, DragonFly BSD 4.4 without ICU