Re: [HACKERS] Lock partitions
Tom Lane wrote: Bruce Momjian [EMAIL PROTECTED] writes: I see this in the CVS commits for 8.2. Did we determine the proper number of lock partitions? Should it be based on the number of buffers or concurrent sessions allowed? No. NUM_LOCK_PARTITIONS needs to be a compile-time constant for a number of reasons, and there is absolutely zero evidence to justify making any effort (and spending any cycles) on a variable value. It would be nice to see some results from the OSDL tests with, say, 4, 8, and 16 lock partitions before we forget about the point though. Anybody know whether OSDL is in a position to run tests for us? I have a couple of bigger runs now against a CVS checkout on 2006-09-20 (please forgive my NUM_BUFFER_PARTITIONS note if you notice that on the web pages): Baseline (default NUM_LOCK_PARTITIONS=4): notpm 6589 http://dbt.osdl.org/dbt/dbt2dev/results/dev4-015/184/ NUM_LOCK_PARTITIONS=8: notpm 4471 http://dbt.osdl.org/dbt/dbt2dev/results/dev4-015/180/ NUM_LOCK_PARTITIONS=16: Failed to run. The number of transaction errors increased when I increased the NUM_LOCK_PARTITIONS, which I think is the reason it failed to run when I set it to 16. And the throughput went down significantly (32%). Should I try against a more recent checkout? Mark ---(end of broadcast)--- TIP 1: if posting/reading through Usenet, please send an appropriate subscribe-nomail command to [EMAIL PROTECTED] so that your message can get through to the mailing list cleanly
Re: [HACKERS] Lock partitions
Mark Wong [EMAIL PROTECTED] writes: The number of transaction errors increased when I increased the NUM_LOCK_PARTITIONS, which I think is the reason it failed to run when I set it to 16. Hmm, what sort of errors are we talking about? I wonder if you've exposed a bug. Changing NUM_LOCK_PARTITIONS really shouldn't have any semantic effect. regards, tom lane ---(end of broadcast)--- TIP 2: Don't 'kill -9' the postmaster
Re: [HACKERS] Lock partitions
Tom Lane wrote: Mark Wong [EMAIL PROTECTED] writes: The number of transaction errors increased when I increased the NUM_LOCK_PARTITIONS, which I think is the reason it failed to run when I set it to 16. Hmm, what sort of errors are we talking about? I wonder if you've exposed a bug. Changing NUM_LOCK_PARTITIONS really shouldn't have any semantic effect. The libpq client (error log: http://dbt.osdl.org/dbt/dbt2dev/results/dev4-015/180/client/error.log) is saying things like this: ERROR: too many LWLocks taken CONTEXT: SQL statement DELETE FROM new_order WHERE no_o_id = 2101 AND no_w_id = 349 AND no_d_id = 1 A grep through that file shows that all the unexpected errors appear to be due to too many LWLocks taken. Mark ---(end of broadcast)--- TIP 2: Don't 'kill -9' the postmaster
Re: [HACKERS] Lock partitions
Mark Wong [EMAIL PROTECTED] writes: Tom Lane wrote: Hmm, what sort of errors are we talking about? ERROR: too many LWLocks taken That really shouldn't happen ... are you sure you did a full recompile after changing NUM_LOCK_PARTITIONS? Actually ... wait a moment. The default value of NUM_LOCK_PARTITIONS is already 16 (1 LOG2_NUM_LOCK_PARTITIONS where the latter is 4). Are you saying you set LOG2_NUM_LOCK_PARTITIONS to 16? That would be way too many partitions. I was thinking of testing LOG2_NUM_LOCK_PARTITIONS in the range of about 2 to 5. regards, tom lane ---(end of broadcast)--- TIP 6: explain analyze is your friend
Re: [HACKERS] Lock partitions
Tom Lane wrote: Mark Wong [EMAIL PROTECTED] writes: Tom Lane wrote: Hmm, what sort of errors are we talking about? ERROR: too many LWLocks taken That really shouldn't happen ... are you sure you did a full recompile after changing NUM_LOCK_PARTITIONS? Actually ... wait a moment. The default value of NUM_LOCK_PARTITIONS is already 16 (1 LOG2_NUM_LOCK_PARTITIONS where the latter is 4). Are you saying you set LOG2_NUM_LOCK_PARTITIONS to 16? That would be way too many partitions. I was thinking of testing LOG2_NUM_LOCK_PARTITIONS in the range of about 2 to 5. Oops, I can't read bit shifting. =p I'll do again. Mark ---(end of broadcast)--- TIP 3: Have you checked our extensive FAQ? http://www.postgresql.org/docs/faq
Re: [HACKERS] Lock partitions
Mark Wong wrote:
Tom Lane wrote:
Mark Wong [EMAIL PROTECTED] writes:
Curious, I'm still seeing the same behavior. Maybe I'll take another
snapshot from CVS.
Hm, maybe I need to try a bit harder here. Does the not registered
error happen immediately/reliably for you, or do you need to run the
test awhile?
I did a gross test and my kit appears broken between the 8.0 and 8.1
releases. I'll try to narrow down the exact date.
I've narrowed it down between cvs pulls from Dec 14, 2005 and Dec 15,
2005. Does the attached diff appear to be a plausible cause?
Thanks,
Mark
diff -urN pgsql-2005-12-14/src/backend/commands/prepare.c
pgsql-2005-12-15/src/backend/commands/prepare.c
--- pgsql-2005-12-14/src/backend/commands/prepare.c 2005-11-28
17:25:49.0 -0800
+++ pgsql-2005-12-15/src/backend/commands/prepare.c 2005-12-14
09:06:27.0 -0800
@@ -10,7 +10,7 @@
* Copyright (c) 2002-2005, PostgreSQL Global Development Group
*
* IDENTIFICATION
- * $PostgreSQL: pgsql/src/backend/commands/prepare.c,v 1.43 2005/11/29
01:25:49 tgl Exp $
+ * $PostgreSQL: pgsql/src/backend/commands/prepare.c,v 1.44 2005/12/14
17:06:27 tgl Exp $
*
*-
*/
@@ -448,6 +448,30 @@
}
/*
+ * Given a prepared statement, determine whether it will return tuples.
+ *
+ * Note: this is used rather than just testing the result of
+ * FetchPreparedStatementResultDesc() because that routine can fail if
+ * invoked in an aborted transaction. This one is safe to use in any
+ * context. Be sure to keep the two routines in sync!
+ */
+bool
+PreparedStatementReturnsTuples(PreparedStatement *stmt)
+{
+ switch (ChoosePortalStrategy(stmt-query_list))
+ {
+ case PORTAL_ONE_SELECT:
+ case PORTAL_UTIL_SELECT:
+ return true;
+
+ case PORTAL_MULTI_QUERY:
+ /* will not return tuples */
+ break;
+ }
+ return false;
+}
+
+/*
* Given a prepared statement that returns tuples, extract the query
* targetlist. Returns NIL if the statement doesn't have a determinable
* targetlist.
diff -urN pgsql-2005-12-14/src/backend/executor/execQual.c
pgsql-2005-12-15/src/backend/executor/execQual.c
--- pgsql-2005-12-14/src/backend/executor/execQual.c2005-11-22
10:17:10.0 -0800
+++ pgsql-2005-12-15/src/backend/executor/execQual.c2005-12-14
08:28:32.0 -0800
@@ -8,7 +8,7 @@
*
*
* IDENTIFICATION
- * $PostgreSQL: pgsql/src/backend/executor/execQual.c,v 1.185 2005/11/22
18:17:10 momjian Exp $
+ * $PostgreSQL: pgsql/src/backend/executor/execQual.c,v 1.186 2005/12/14
16:28:32 tgl Exp $
*
*-
*/
@@ -523,7 +523,7 @@
Assert(variable-varno != OUTER);
slot = econtext-ecxt_scantuple;
- tuple = slot-tts_tuple;
+ tuple = ExecFetchSlotTuple(slot);
tupleDesc = slot-tts_tupleDescriptor;
/*
diff -urN pgsql-2005-12-14/src/backend/tcop/postgres.c
pgsql-2005-12-15/src/backend/tcop/postgres.c
--- pgsql-2005-12-14/src/backend/tcop/postgres.c2005-11-22
10:17:21.0 -0800
+++ pgsql-2005-12-15/src/backend/tcop/postgres.c2005-12-14
09:06:27.0 -0800
@@ -8,7 +8,7 @@
*
*
* IDENTIFICATION
- * $PostgreSQL: pgsql/src/backend/tcop/postgres.c,v 1.470 2005/11/22
18:17:21 momjian Exp $
+ * $PostgreSQL: pgsql/src/backend/tcop/postgres.c,v 1.471 2005/12/14
17:06:27 tgl Exp $
*
* NOTES
* this is the main module of the postgres backend and
@@ -1849,6 +1849,15 @@
ListCell *l;
StringInfoData buf;
+ /*
+* Start up a transaction command. (Note that this will normally change
+* current memory context.) Nothing happens if we are already in one.
+*/
+ start_xact_command();
+
+ /* Switch back to message context */
+ MemoryContextSwitchTo(MessageContext);
+
/* Find prepared statement */
if (stmt_name[0] != '\0')
pstmt = FetchPreparedStatement(stmt_name, true);
@@ -1862,6 +1871,22 @@
errmsg(unnamed prepared statement
does not exist)));
}
+ /*
+* If we are in aborted transaction state, we can't safely create a
result
+* tupledesc, because that needs catalog accesses. Hence, refuse to
+* Describe statements that return data. (We shouldn't just refuse all
+* Describes, since that might break the ability of some clients to
issue
+* COMMIT or ROLLBACK commands, if they use code that blindly Describes
+* whatever it does.) We can Describe parameters without doing anything
+* dangerous, so we don't restrict that.
+*/
+ if (IsAbortedTransactionBlockState()
+ PreparedStatementReturnsTuples(pstmt))
+
Re: [HACKERS] Lock partitions
Mark Wong [EMAIL PROTECTED] writes: I did a gross test and my kit appears broken between the 8.0 and 8.1 releases. I'll try to narrow down the exact date. I've narrowed it down between cvs pulls from Dec 14, 2005 and Dec 15, 2005. Does the attached diff appear to be a plausible cause? No, not particularly. You sure about those dates? regards, tom lane ---(end of broadcast)--- TIP 3: Have you checked our extensive FAQ? http://www.postgresql.org/docs/faq
Re: [HACKERS] Lock partitions
Tom Lane wrote: Mark Wong [EMAIL PROTECTED] writes: I did a gross test and my kit appears broken between the 8.0 and 8.1 releases. I'll try to narrow down the exact date. I've narrowed it down between cvs pulls from Dec 14, 2005 and Dec 15, 2005. Does the attached diff appear to be a plausible cause? No, not particularly. You sure about those dates? Ugh, double checking tells me I'm wrong. I'll keep testing. Thanks, Mark ---(end of broadcast)--- TIP 2: Don't 'kill -9' the postmaster
Re: [HACKERS] Lock partitions
Tom Lane wrote: Mark Wong [EMAIL PROTECTED] writes: Curious, I'm still seeing the same behavior. Maybe I'll take another snapshot from CVS. Hm, maybe I need to try a bit harder here. Does the not registered error happen immediately/reliably for you, or do you need to run the test awhile? I did a gross test and my kit appears broken between the 8.0 and 8.1 releases. I'll try to narrow down the exact date. Mark ---(end of broadcast)--- TIP 2: Don't 'kill -9' the postmaster
Re: [HACKERS] Lock partitions
Tom, Taking the 4 lock vs 8 lock partitions, 4 LockMgr lock partitions spent a total of 652 seconds in lock management (acquiring/releasing) and 8 LockMgr lock partitions spent a total of 536 in lock management. This is an improvement of 116 seconds, but the TPS didn't improve by much - only a 1.21 TPS improvement. The improvement in the LockMgr processing is consumed by the next system bottleneck downstream as more work is being let through. In this particular case it's the WALInsertLock lock. The 4 LockMgr lock partition test spent a total of 5868 seconds in WALInsertLock lock management whereas the 8 LockMgr partition test spent 5945 seconds in WALInsertLock lock management which is an increase of 77 seconds. But, that's not the only static lock that increased in time, it's just the most significant increase. The WALWriteLock lock increased by 12 seconds, ProcArrayLock increased by 8 seconds and SInvalLock increased by 5 seconds. This takes the total time flowing to other locks to 102 seconds. The locks are not the only part of the puzzle. As improvements are made to various areas like the BufMapping and LockMgr lock partitions, other parts of the system start to get exercised in ways that were not possible in previous releases. We're still trying to get our arms around all the functions that might become bottlenecks when other lock contention is minimized. And, improvements are being made. The locking changes from 8.0.x to 8.1.x made a significant difference in scalability. Again, the current lock improvements in 8.2 have realized ~20% improvement over 8.1.x, based on our testing. We added monitoring code to the LWLockAcquire and LWLockRelease functions. We track the total time taken to pass through LWLockAcquire and LWLockRelease. So, if a particular backend process takes 1 second to run through LWLockAcquire, we will track that as 1 second in lock acquisition. Irrespective of whether my backend process was spinning or in a semaphore wait, it's 1 second that was taken away from processing a statement/request. We could also add timing for semaphore waits within LWLockAcquire, if that would be a useful statistic. Let me know if there are any other tests or metrics that would be useful. David -Original Message- From: Tom Lane [mailto:[EMAIL PROTECTED] Sent: Wednesday, September 13, 2006 1:36 PM To: Strong, David Cc: PostgreSQL-development Subject: Re: [HACKERS] Lock partitions Strong, David [EMAIL PROTECTED] writes: We have some results for you. We left the buffer partition locks at 128 as this did not seem to be a concern and we're still using 25 backend processes. We ran tests for 4, 8 and 16 lock partitions. For 4 lock partitions, it took 620 seconds to acquire locks and 32 seconds to release locks. The test produced 199.95 TPS. For 8 lock partitions, it took 505 seconds to acquire locks and 31 seconds to release locks. The test produced 201.16 TPS. For 16 lock partitions, it took 362 seconds to acquire locks and 22 seconds to release locks. The test produced 200.75 TPS. And, just for grins, using 128 buffer and 128 lock partitions, took 235 seconds to acquire locks and 22 seconds to release locks. The test produced 203.24 TPS. [ itch... ] I can't help thinking there's something wrong with this; the wait-time measurements seem sane, but why is there essentially no change in the TPS result? The above numbers are only for the lock-partition LWLocks, right? What are the totals --- that is, how much time is spent blocked vs. processing overall? regards, tom lane ---(end of broadcast)--- TIP 3: Have you checked our extensive FAQ? http://www.postgresql.org/docs/faq
Re: [HACKERS] Lock partitions
Tom Lane wrote:
Mark Wong [EMAIL PROTECTED] writes:
Tom Lane wrote:
It would be nice to see some results from the OSDL tests with, say, 4,
8, and 16 lock partitions before we forget about the point though.
Anybody know whether OSDL is in a position to run tests for us?
Yeah, I can run some dbt2 tests in the lab. I'll get started on it.
We're still a little bit away from getting the automated testing for
PostgreSQL going again though.
Great, thanks. The thing to twiddle is LOG2_NUM_LOCK_PARTITIONS in
src/include/storage/lwlock.h. You need a full backend recompile
after changing it, but you shouldn't need to initdb, if that helps.
Sorry for the delay but looks like there's some data coming in. It also
looks like my kit is starting to be a little dated. My stored libpq
calls are failing. I'm getting this message:
ERROR: record type has not been registered
From PQerrorMessage() on line 41 from this bit of code:
/* Create the query and execute it. */
sprintf(stmt, SELECT * FROM order_status(%d, %d, %d, '%s'),
data-c_id, data-c_w_id, data-c_d_id, data-c_last);
res = PQexec(dbc-conn, stmt);
if (!res || (PQresultStatus(res) != PGRES_COMMAND_OK
PQresultStatus(res) != PGRES_TUPLES_OK)) {
LOG_ERROR_MESSAGE(%s, PQerrorMessage(dbc-conn));
PQclear(res);
return ERROR;
}
LOG_ERROR_MESSAGE() is just a macro for a function that does a printf().
Any suggestions?
Thanks,
Mark
---(end of broadcast)---
TIP 6: explain analyze is your friend
Re: [HACKERS] Lock partitions
Mark Wong [EMAIL PROTECTED] writes: Sorry for the delay but looks like there's some data coming in. It also looks like my kit is starting to be a little dated. My stored libpq calls are failing. I'm getting this message: ERROR: record type has not been registered This is a server-side failure --- could we see how order_status() is defined? What PG version are you testing exactly? regards, tom lane ---(end of broadcast)--- TIP 2: Don't 'kill -9' the postmaster
Re: [HACKERS] Lock partitions
Tom Lane wrote: Mark Wong [EMAIL PROTECTED] writes: Sorry for the delay but looks like there's some data coming in. It also looks like my kit is starting to be a little dated. My stored libpq calls are failing. I'm getting this message: ERROR: record type has not been registered This is a server-side failure --- could we see how order_status() is defined? What PG version are you testing exactly? I took pgsqsl snapshot from cvs on Sept 11. Due to the length of the file that order_status() is in and of order_status() itself, here's is a url for the file in the svn repository. order_status() is defined starting on line 710: http://svn.sourceforge.net/viewvc/osdldbt/trunk/dbt2/storedproc/pgsql/c/funcs.c?view=markup Thanks, Mark ---(end of broadcast)--- TIP 4: Have you searched our list archives? http://archives.postgresql.org
Re: [HACKERS] Lock partitions
Mark Wong [EMAIL PROTECTED] writes: Tom Lane wrote: This is a server-side failure --- could we see how order_status() is defined? What PG version are you testing exactly? I took pgsqsl snapshot from cvs on Sept 11. Due to the length of the file that order_status() is in and of order_status() itself, here's is a url for the file in the svn repository. order_status() is defined starting on line 710: http://svn.sourceforge.net/viewvc/osdldbt/trunk/dbt2/storedproc/pgsql/c/funcs.c?view=markup Hmph. I think we broke something --- the error implies that some function tried to return a tuple that hadn't been properly blessed, but I can't see that order_status or any of the other functions in that file are doing anything wrong. In any case, if it used to work for you, we had better figure out exactly why it stopped working. I know you've posted info before on how to set up and run the dbt code, but could you refresh my memory? Is there a URL somewhere with the info? regards, tom lane ---(end of broadcast)--- TIP 3: Have you checked our extensive FAQ? http://www.postgresql.org/docs/faq
Re: [HACKERS] Lock partitions
Tom Lane wrote: Mark Wong [EMAIL PROTECTED] writes: Tom Lane wrote: This is a server-side failure --- could we see how order_status() is defined? What PG version are you testing exactly? I took pgsqsl snapshot from cvs on Sept 11. Due to the length of the file that order_status() is in and of order_status() itself, here's is a url for the file in the svn repository. order_status() is defined starting on line 710: http://svn.sourceforge.net/viewvc/osdldbt/trunk/dbt2/storedproc/pgsql/c/funcs.c?view=markup Hmph. I think we broke something --- the error implies that some function tried to return a tuple that hadn't been properly blessed, but I can't see that order_status or any of the other functions in that file are doing anything wrong. In any case, if it used to work for you, we had better figure out exactly why it stopped working. I know you've posted info before on how to set up and run the dbt code, but could you refresh my memory? Is there a URL somewhere with the info? Here's a readme: http://svn.sourceforge.net/viewvc/osdldbt/trunk/dbt2/README-POSTGRESQL?view=markup But perhaps something much easier, using subversion: mkdir /mnt/dbt2 # for pgdata svn co https://svn.sourceforge.net/svnroot/osdldbt/trunk/dbt2 dbt2 cd dbt2 ./configure --with-postgresql=pgsql_dir make cd scripts/pgsql/ ./build_db -g -w 1 cd .. ./run_workload -w 1 -d 120 -c 20 -n I think that should work. That will create a 1 warehouse database and run a 120 second test with no-thinktimes. That should be sufficient to run through each transaction more than several times. Thanks, Mark ---(end of broadcast)--- TIP 3: Have you checked our extensive FAQ? http://www.postgresql.org/docs/faq
Re: [HACKERS] Lock partitions
Mark Wong [EMAIL PROTECTED] writes: But perhaps something much easier, using subversion: mkdir /mnt/dbt2 # for pgdata svn co https://svn.sourceforge.net/svnroot/osdldbt/trunk/dbt2 dbt2 cd dbt2 ./configure --with-postgresql=pgsql_dir configure is not in the svn checkout. I guessed that I needed to do aclocal/automake/autoconf, but automake fails: $ automake configure.ac:11: required file `config.h.in' not found $ and I don't see anyplace to get config.h.in from. regards, tom lane ---(end of broadcast)--- TIP 1: if posting/reading through Usenet, please send an appropriate subscribe-nomail command to [EMAIL PROTECTED] so that your message can get through to the mailing list cleanly
Re: [HACKERS] Lock partitions
Tom Lane wrote: Mark Wong [EMAIL PROTECTED] writes: But perhaps something much easier, using subversion: mkdir /mnt/dbt2 # for pgdata svn co https://svn.sourceforge.net/svnroot/osdldbt/trunk/dbt2 dbt2 cd dbt2 ./configure --with-postgresql=pgsql_dir configure is not in the svn checkout. I guessed that I needed to do aclocal/automake/autoconf, but automake fails: $ automake configure.ac:11: required file `config.h.in' not found $ and I don't see anyplace to get config.h.in from. Oops! 'autoreconf --install' is what I run to generate all that stuff. Sorry about that. Thanks, Mark ---(end of broadcast)--- TIP 1: if posting/reading through Usenet, please send an appropriate subscribe-nomail command to [EMAIL PROTECTED] so that your message can get through to the mailing list cleanly
Re: [HACKERS] Lock partitions
Mark Wong [EMAIL PROTECTED] writes: Oops! 'autoreconf --install' is what I run to generate all that stuff. Ah, better. I see at least part of the problem: CREATE OR REPLACE FUNCTION stock_level (INTEGER, INTEGER, INTEGER) RETURNS INTEGER AS '/home/tgl/dbt2/storedproc/pgsql/c/../../../storedproc/pgsql/c/funcs' LANGUAGE C STRICT; psql:/home/tgl/dbt2/scripts/pgsql/../../storedproc/pgsql/c/stock_level.sql:7: ERROR: incompatible library /home/tgl/dbt2/storedproc/pgsql/c/../../../storedproc/pgsql/c/funcs.so: missing magic block HINT: Extension libraries are now required to use the PG_MODULE_MAGIC macro. You need to add something like this to funcs.c: #include executor/spi.h /* this should include most necessary APIs */ #include executor/executor.h /* for GetAttributeByName() */ #include funcapi.h /* for returning set of rows in order_status */ + + #ifdef PG_MODULE_MAGIC + PG_MODULE_MAGIC; + #endif /* #define DEBUG With that change, I didn't see run_workload report any errors, but maybe I don't know where to look. I'm not sure how this bug could have led to a type not registered error ... the query should've just failed outright. regards, tom lane ---(end of broadcast)--- TIP 2: Don't 'kill -9' the postmaster
Re: [HACKERS] Lock partitions
Tom Lane wrote: Mark Wong [EMAIL PROTECTED] writes: Oops! 'autoreconf --install' is what I run to generate all that stuff. Ah, better. I see at least part of the problem: CREATE OR REPLACE FUNCTION stock_level (INTEGER, INTEGER, INTEGER) RETURNS INTEGER AS '/home/tgl/dbt2/storedproc/pgsql/c/../../../storedproc/pgsql/c/funcs' LANGUAGE C STRICT; psql:/home/tgl/dbt2/scripts/pgsql/../../storedproc/pgsql/c/stock_level.sql:7: ERROR: incompatible library /home/tgl/dbt2/storedproc/pgsql/c/../../../storedproc/pgsql/c/funcs.so: missing magic block HINT: Extension libraries are now required to use the PG_MODULE_MAGIC macro. You need to add something like this to funcs.c: #include executor/spi.h /* this should include most necessary APIs */ #include executor/executor.h /* for GetAttributeByName() */ #include funcapi.h /* for returning set of rows in order_status */ + + #ifdef PG_MODULE_MAGIC + PG_MODULE_MAGIC; + #endif /* #define DEBUG With that change, I didn't see run_workload report any errors, but maybe I don't know where to look. I'm not sure how this bug could have led to a type not registered error ... the query should've just failed outright. The error is captured in dbt2/scripts/output/*/client/error.log, where * is the run directory. Ok, I'll give it a shot on my system. Thanks, Mark ---(end of broadcast)--- TIP 6: explain analyze is your friend
Re: [HACKERS] Lock partitions
Mark Wong [EMAIL PROTECTED] writes: Tom Lane wrote: With that change, I didn't see run_workload report any errors, but maybe I don't know where to look. The error is captured in dbt2/scripts/output/*/client/error.log, where * is the run directory. Hm ... here's what I see in there: Thu Sep 14 15:19:16 2006 tid:-1430387232 client.c:129 20 DB worker threads have started Thu Sep 14 15:19:31 2006 tid:1087957312 libpq/dbc_new_order.c:111 ERROR: deadlock detected DETAIL: Process 5334 waits for ShareLock on transaction 3505055; blocked by process 5363. Process 5363 waits for ShareLock on transaction 3505049; blocked by process 5334. CONTEXT: SQL statement UPDATE stock SET s_quantity = s_quantity - 10 WHERE s_i_id = 48368 AND s_w_id = 1 Thu Sep 14 15:21:18 2006 tid:1089960256 client_interface.c:33 socket closed on _receive Thu Sep 14 15:21:18 2006 tid:1089689920 client_interface.c:33 socket closed on _receive Thu Sep 14 15:21:18 2006 tid:1090636096 client_interface.c:33 socket closed on _receive Thu Sep 14 15:21:18 2006 tid:1090230592 client_interface.c:33 socket closed on _receive Thu Sep 14 15:21:18 2006 tid:1090365760 client_interface.c:33 socket closed on _receive Thu Sep 14 15:21:18 2006 tid:1090095424 client_interface.c:33 socket closed on _receive Thu Sep 14 15:21:18 2006 tid:1089825088 client_interface.c:33 socket closed on _receive Thu Sep 14 15:21:18 2006 tid:1090500928 client_interface.c:33 socket closed on _receive Thu Sep 14 15:21:18 2006 tid:1090906432 client_interface.c:33 socket closed on _receive Thu Sep 14 15:21:18 2006 tid:1090771264 client_interface.c:33 socket closed on _receive Is the deadlock failure expected? regards, tom lane ---(end of broadcast)--- TIP 3: Have you checked our extensive FAQ? http://www.postgresql.org/docs/faq
Re: [HACKERS] Lock partitions
Tom Lane wrote: Mark Wong [EMAIL PROTECTED] writes: Tom Lane wrote: With that change, I didn't see run_workload report any errors, but maybe I don't know where to look. The error is captured in dbt2/scripts/output/*/client/error.log, where * is the run directory. Hm ... here's what I see in there: Thu Sep 14 15:19:16 2006 tid:-1430387232 client.c:129 20 DB worker threads have started Thu Sep 14 15:19:31 2006 tid:1087957312 libpq/dbc_new_order.c:111 ERROR: deadlock detected DETAIL: Process 5334 waits for ShareLock on transaction 3505055; blocked by process 5363. Process 5363 waits for ShareLock on transaction 3505049; blocked by process 5334. CONTEXT: SQL statement UPDATE stock SET s_quantity = s_quantity - 10 WHERE s_i_id = 48368 AND s_w_id = 1 [snip] Is the deadlock failure expected? Ooh, that's interesting. Deadlock failure is possible although I think we would all prefer that it didn't happen. In the scheme of the workload having failed transactions is ok. So with respect to having an invalid test run it's something I wouldn't worry about too much if it's infrequent. Mark ---(end of broadcast)--- TIP 3: Have you checked our extensive FAQ? http://www.postgresql.org/docs/faq
Re: [HACKERS] Lock partitions
Mark Wong [EMAIL PROTECTED] writes: Curious, I'm still seeing the same behavior. Maybe I'll take another snapshot from CVS. Hm, maybe I need to try a bit harder here. Does the not registered error happen immediately/reliably for you, or do you need to run the test awhile? As for the deadlock issue you mention I've been told I have some seeding/random number generation problems in the kit. Perhaps that is related to the deadlock at least. The nature of the locks suggests that you have two transactions trying to update the same two rows in opposite orders. The usually recommended fix is to ensure you use a consistent processing order within a transaction (eg, ascending primary keys --- but any well-defined row ordering will work). regards, tom lane ---(end of broadcast)--- TIP 9: In versions below 8.0, the planner will ignore your desire to choose an index scan if your joining column's datatypes do not match
Re: [HACKERS] Lock partitions
Mark Wong [EMAIL PROTECTED] writes: Curious, I'm still seeing the same behavior. Maybe I'll take another snapshot from CVS. Hm, maybe I need to try a bit harder here. Does the not registered error happen immediately/reliably for you, or do you need to run the test awhile? It appears to occur for every order status transaction. You can monitor the transactions by watching dbt2/scripts/output/*/driver/mix.log. A 'o' (lowercase) indicates a successful order status transaction while a 'O' (uppercase) indiciates an unsuccessful transaction. or, perhaps a simplier way is to start the database then: cd dbt2/src ./transaction_test -d localhost -l 5432 -t o That will connect directly to the database and execute an order status transaction. Thanks, Mark ---(end of broadcast)--- TIP 4: Have you searched our list archives? http://archives.postgresql.org
Re: [HACKERS] Lock partitions
On Tue, 2006-09-12 at 12:40 -0400, Tom Lane wrote: Strong, David [EMAIL PROTECTED] writes: When using 16 buffer and 16 lock partitions, we see that BufMapping takes 809 seconds to acquire locks and 174 seconds to release locks. The LockMgr takes 362 seconds to acquire locks and 26 seconds to release locks. When using 128 buffer and 128 lock partitions, we see that BufMapping takes 277 seconds (532 seconds improvement) to acquire locks and 78 seconds (96 seconds improvement) to release locks. The LockMgr takes 235 seconds (127 seconds improvement) to acquire locks and 22 seconds (4 seconds improvement) to release locks. While I don't see any particular penalty to increasing NUM_BUFFER_PARTITIONS, increasing NUM_LOCK_PARTITIONS carries a very significant penalty (increasing PGPROC size as well as the work needed during LockReleaseAll, which is executed at every transaction end). I think 128 lock partitions is probably verging on the ridiculous ... particularly if your benchmark only involves touching half a dozen tables. I'd be more interested in comparisons between 4 and 16 lock partitions. Also, please vary the two settings independently rather than confusing the issue by changing them both at once. Good thinking David. Even if 128 is fairly high, it does seem worth exploring higher values - I was just stuck in fewer == better thoughts. With the improvements in the various locking times, one might expect an improvement in the overall benchmark result. However, a 16 partition run produces a result of 198.74 TPS and a 128 partition run produces a result of 203.24 TPS. Part of the time saved from BufMapping and LockMgr partitions is absorbed into the WALInsertLock lock. For a 16 partition run, the total time to lock/release the WALInsertLock lock is 5845 seconds. For 128 partitions, the WALInsertLock lock takes 6172 seconds, an increase of 327 seconds. Perhaps we have our WAL configured incorrectly? I fear this throws your entire measurement procedure into question. For a fixed workload the number of acquisitions of WALInsertLock ought to be fixed, so you shouldn't see any more contention for WALInsertLock if the transaction rate didn't change materially. David's results were to do with lock acquire/release time, not the number of acquisitions, so that in itself doesn't make me doubt these measurements. Perhaps we can ask whether there was a substantially different number of lock acquisitions? As Tom says, that would be an issue. It seems reasonable that relieving the bottleneck on BufMapping and LockMgr locks that we would then queue longer on the next bottleneck, WALInsertLock. So again, those tests seem reasonable to me so far. These seem to be the beginnings of accurate wait time analysis, so I'm listening closely. Are you using a lightweight timer? -- Simon Riggs EnterpriseDB http://www.enterprisedb.com ---(end of broadcast)--- TIP 3: Have you checked our extensive FAQ? http://www.postgresql.org/docs/faq
Re: [HACKERS] Lock partitions
Simon, In the 16/16 (16 buffer partitions/16 lock partitions) test, the WALInsertLock lock had 14643080 acquisition attempts and 12057678 successful acquisitions on the lock. That's 2585402 retries on the lock. That is to say that PGSemaphoreLock was invoked 2585402 times. In the 128/128 test, the WALInsertLock lock had 14991208 acquisition attempts and 12324765 successful acquisitions. That's 2666443 retries. The 128/128 test attempted 348128 more lock acquisitions than the 16/16 test and retried 81041 times more than the 16/16 test. We attribute the rise in WALInsertLock lock accesses to the reduction in time on acquiring the BufMapping and LockMgr partition locks. Does this seem reasonable? The overhead of any monitoring is of great concern to us. We've tried both clock_gettime () and getttimeofday () calls. They both seem to have the same overhead ~1 us/call (measured against the TSC of the CPU) and both seem to be accurate. We realize this can be a delicate point and so we would be happy to rerun any tests with a different timing mechanism. David -Original Message- From: Simon Riggs [mailto:[EMAIL PROTECTED] Sent: Wednesday, September 13, 2006 2:22 AM To: Tom Lane Cc: Strong, David; PostgreSQL-development Subject: Re: [HACKERS] Lock partitions On Tue, 2006-09-12 at 12:40 -0400, Tom Lane wrote: Strong, David [EMAIL PROTECTED] writes: When using 16 buffer and 16 lock partitions, we see that BufMapping takes 809 seconds to acquire locks and 174 seconds to release locks. The LockMgr takes 362 seconds to acquire locks and 26 seconds to release locks. When using 128 buffer and 128 lock partitions, we see that BufMapping takes 277 seconds (532 seconds improvement) to acquire locks and 78 seconds (96 seconds improvement) to release locks. The LockMgr takes 235 seconds (127 seconds improvement) to acquire locks and 22 seconds (4 seconds improvement) to release locks. While I don't see any particular penalty to increasing NUM_BUFFER_PARTITIONS, increasing NUM_LOCK_PARTITIONS carries a very significant penalty (increasing PGPROC size as well as the work needed during LockReleaseAll, which is executed at every transaction end). I think 128 lock partitions is probably verging on the ridiculous ... particularly if your benchmark only involves touching half a dozen tables. I'd be more interested in comparisons between 4 and 16 lock partitions. Also, please vary the two settings independently rather than confusing the issue by changing them both at once. Good thinking David. Even if 128 is fairly high, it does seem worth exploring higher values - I was just stuck in fewer == better thoughts. With the improvements in the various locking times, one might expect an improvement in the overall benchmark result. However, a 16 partition run produces a result of 198.74 TPS and a 128 partition run produces a result of 203.24 TPS. Part of the time saved from BufMapping and LockMgr partitions is absorbed into the WALInsertLock lock. For a 16 partition run, the total time to lock/release the WALInsertLock lock is 5845 seconds. For 128 partitions, the WALInsertLock lock takes 6172 seconds, an increase of 327 seconds. Perhaps we have our WAL configured incorrectly? I fear this throws your entire measurement procedure into question. For a fixed workload the number of acquisitions of WALInsertLock ought to be fixed, so you shouldn't see any more contention for WALInsertLock if the transaction rate didn't change materially. David's results were to do with lock acquire/release time, not the number of acquisitions, so that in itself doesn't make me doubt these measurements. Perhaps we can ask whether there was a substantially different number of lock acquisitions? As Tom says, that would be an issue. It seems reasonable that relieving the bottleneck on BufMapping and LockMgr locks that we would then queue longer on the next bottleneck, WALInsertLock. So again, those tests seem reasonable to me so far. These seem to be the beginnings of accurate wait time analysis, so I'm listening closely. Are you using a lightweight timer? -- Simon Riggs EnterpriseDB http://www.enterprisedb.com ---(end of broadcast)--- TIP 9: In versions below 8.0, the planner will ignore your desire to choose an index scan if your joining column's datatypes do not match
Re: [HACKERS] Lock partitions
Tom, We have some results for you. We left the buffer partition locks at 128 as this did not seem to be a concern and we're still using 25 backend processes. We ran tests for 4, 8 and 16 lock partitions. For 4 lock partitions, it took 620 seconds to acquire locks and 32 seconds to release locks. The test produced 199.95 TPS. For 8 lock partitions, it took 505 seconds to acquire locks and 31 seconds to release locks. The test produced 201.16 TPS. For 16 lock partitions, it took 362 seconds to acquire locks and 22 seconds to release locks. The test produced 200.75 TPS. And, just for grins, using 128 buffer and 128 lock partitions, took 235 seconds to acquire locks and 22 seconds to release locks. The test produced 203.24 TPS. Let me know if we can provide any additional information from these tests and if there are any other tests that we can run. David -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of Strong, David Sent: Wednesday, September 13, 2006 10:52 AM To: PostgreSQL-development Subject: Re: [HACKERS] Lock partitions Simon, In the 16/16 (16 buffer partitions/16 lock partitions) test, the WALInsertLock lock had 14643080 acquisition attempts and 12057678 successful acquisitions on the lock. That's 2585402 retries on the lock. That is to say that PGSemaphoreLock was invoked 2585402 times. In the 128/128 test, the WALInsertLock lock had 14991208 acquisition attempts and 12324765 successful acquisitions. That's 2666443 retries. The 128/128 test attempted 348128 more lock acquisitions than the 16/16 test and retried 81041 times more than the 16/16 test. We attribute the rise in WALInsertLock lock accesses to the reduction in time on acquiring the BufMapping and LockMgr partition locks. Does this seem reasonable? The overhead of any monitoring is of great concern to us. We've tried both clock_gettime () and getttimeofday () calls. They both seem to have the same overhead ~1 us/call (measured against the TSC of the CPU) and both seem to be accurate. We realize this can be a delicate point and so we would be happy to rerun any tests with a different timing mechanism. David -Original Message- From: Simon Riggs [mailto:[EMAIL PROTECTED] Sent: Wednesday, September 13, 2006 2:22 AM To: Tom Lane Cc: Strong, David; PostgreSQL-development Subject: Re: [HACKERS] Lock partitions On Tue, 2006-09-12 at 12:40 -0400, Tom Lane wrote: Strong, David [EMAIL PROTECTED] writes: When using 16 buffer and 16 lock partitions, we see that BufMapping takes 809 seconds to acquire locks and 174 seconds to release locks. The LockMgr takes 362 seconds to acquire locks and 26 seconds to release locks. When using 128 buffer and 128 lock partitions, we see that BufMapping takes 277 seconds (532 seconds improvement) to acquire locks and 78 seconds (96 seconds improvement) to release locks. The LockMgr takes 235 seconds (127 seconds improvement) to acquire locks and 22 seconds (4 seconds improvement) to release locks. While I don't see any particular penalty to increasing NUM_BUFFER_PARTITIONS, increasing NUM_LOCK_PARTITIONS carries a very significant penalty (increasing PGPROC size as well as the work needed during LockReleaseAll, which is executed at every transaction end). I think 128 lock partitions is probably verging on the ridiculous ... particularly if your benchmark only involves touching half a dozen tables. I'd be more interested in comparisons between 4 and 16 lock partitions. Also, please vary the two settings independently rather than confusing the issue by changing them both at once. Good thinking David. Even if 128 is fairly high, it does seem worth exploring higher values - I was just stuck in fewer == better thoughts. With the improvements in the various locking times, one might expect an improvement in the overall benchmark result. However, a 16 partition run produces a result of 198.74 TPS and a 128 partition run produces a result of 203.24 TPS. Part of the time saved from BufMapping and LockMgr partitions is absorbed into the WALInsertLock lock. For a 16 partition run, the total time to lock/release the WALInsertLock lock is 5845 seconds. For 128 partitions, the WALInsertLock lock takes 6172 seconds, an increase of 327 seconds. Perhaps we have our WAL configured incorrectly? I fear this throws your entire measurement procedure into question. For a fixed workload the number of acquisitions of WALInsertLock ought to be fixed, so you shouldn't see any more contention for WALInsertLock if the transaction rate didn't change materially. David's results were to do with lock acquire/release time, not the number of acquisitions, so that in itself doesn't make me doubt these measurements. Perhaps we can ask whether there was a substantially different number of lock acquisitions? As Tom says, that would be an issue. It seems reasonable that relieving the bottleneck on BufMapping and LockMgr locks that we
Re: [HACKERS] Lock partitions
Strong, David [EMAIL PROTECTED] writes: We have some results for you. We left the buffer partition locks at 128 as this did not seem to be a concern and we're still using 25 backend processes. We ran tests for 4, 8 and 16 lock partitions. For 4 lock partitions, it took 620 seconds to acquire locks and 32 seconds to release locks. The test produced 199.95 TPS. For 8 lock partitions, it took 505 seconds to acquire locks and 31 seconds to release locks. The test produced 201.16 TPS. For 16 lock partitions, it took 362 seconds to acquire locks and 22 seconds to release locks. The test produced 200.75 TPS. And, just for grins, using 128 buffer and 128 lock partitions, took 235 seconds to acquire locks and 22 seconds to release locks. The test produced 203.24 TPS. [ itch... ] I can't help thinking there's something wrong with this; the wait-time measurements seem sane, but why is there essentially no change in the TPS result? The above numbers are only for the lock-partition LWLocks, right? What are the totals --- that is, how much time is spent blocked vs. processing overall? regards, tom lane ---(end of broadcast)--- TIP 4: Have you searched our list archives? http://archives.postgresql.org
Re: [HACKERS] Lock partitions
On Sep 13, 2006, at 2:46 PM, Strong, David wrote: We have some results for you. We left the buffer partition locks at 128 as this did not seem to be a concern and we're still using 25 backend processes. We ran tests for 4, 8 and 16 lock partitions. Isn't having more lock partitions than buffer partitions pointless? -- Jim Nasby[EMAIL PROTECTED] EnterpriseDB http://enterprisedb.com 512.569.9461 (cell) ---(end of broadcast)--- TIP 5: don't forget to increase your free space map settings
Re: [HACKERS] Lock partitions
Jim Nasby [EMAIL PROTECTED] writes: Isn't having more lock partitions than buffer partitions pointless? AFAIK they're pretty orthogonal. It's true though that a typical transaction doesn't hold all that many locks, which is why I don't see a need for a large number of lock partitions. regards, tom lane ---(end of broadcast)--- TIP 3: Have you checked our extensive FAQ? http://www.postgresql.org/docs/faq
Re: [HACKERS] Lock partitions
On Mon, 2006-09-11 at 11:29 -0400, Tom Lane wrote: Mark Wong [EMAIL PROTECTED] writes: Tom Lane wrote: It would be nice to see some results from the OSDL tests with, say, 4, 8, and 16 lock partitions before we forget about the point though. Anybody know whether OSDL is in a position to run tests for us? Yeah, I can run some dbt2 tests in the lab. I'll get started on it. We're still a little bit away from getting the automated testing for PostgreSQL going again though. Great, thanks. The thing to twiddle is LOG2_NUM_LOCK_PARTITIONS in src/include/storage/lwlock.h. You need a full backend recompile after changing it, but you shouldn't need to initdb, if that helps. IIRC we did that already and the answer was 16... -- Simon Riggs EnterpriseDB http://www.enterprisedb.com ---(end of broadcast)--- TIP 6: explain analyze is your friend
Re: [HACKERS] Lock partitions
Simon Riggs [EMAIL PROTECTED] writes: On Mon, 2006-09-11 at 11:29 -0400, Tom Lane wrote: Great, thanks. The thing to twiddle is LOG2_NUM_LOCK_PARTITIONS in src/include/storage/lwlock.h. You need a full backend recompile after changing it, but you shouldn't need to initdb, if that helps. IIRC we did that already and the answer was 16... No, no one has shown me any numbers from any real tests (anything more than pgbench on a Dell PC ...). regards, tom lane ---(end of broadcast)--- TIP 5: don't forget to increase your free space map settings
Re: [HACKERS] Lock partitions
We can pass on what we've seen when running tests here with different BufMapping and LockMgr partition sizes. We use a TPC-C inspired benchmark. Currently it is configured to run 25 backend processes. The test runs for 16 minutes as this is the minimum amount of time we can run and obtain useful information. This gives us 24,000 seconds (25 * 16 * 60) of processing time. The following timings have been rounded to the nearest second and represent the amount of time amongst all backend processes to acquire and release locks. For example, a value of 2500 seconds would mean each backend process (25) took ~100 seconds to acquire or release a lock. Although, in reality, the time spent locking or releasing each partition entry is not uniform and there are some definite hotspot entries. We can pass on some of the lock output if anyone is interested. When using 16 buffer and 16 lock partitions, we see that BufMapping takes 809 seconds to acquire locks and 174 seconds to release locks. The LockMgr takes 362 seconds to acquire locks and 26 seconds to release locks. When using 128 buffer and 128 lock partitions, we see that BufMapping takes 277 seconds (532 seconds improvement) to acquire locks and 78 seconds (96 seconds improvement) to release locks. The LockMgr takes 235 seconds (127 seconds improvement) to acquire locks and 22 seconds (4 seconds improvement) to release locks. Overall, 128 BufMapping partitions improves locking/releasing by 678 seconds, 128 LockMgr partitions improves locking/releasing by 131 seconds. With the improvements in the various locking times, one might expect an improvement in the overall benchmark result. However, a 16 partition run produces a result of 198.74 TPS and a 128 partition run produces a result of 203.24 TPS. Part of the time saved from BufMapping and LockMgr partitions is absorbed into the WALInsertLock lock. For a 16 partition run, the total time to lock/release the WALInsertLock lock is 5845 seconds. For 128 partitions, the WALInsertLock lock takes 6172 seconds, an increase of 327 seconds. Perhaps we have our WAL configured incorrectly? Other static locks are also affected, but not as much as the WALInsertLock lock. For example, the ProcArrayLock lock increases from 337 seconds to 348 seconds. The SInvalLock lock increases from 317 seconds to 331 seconds. Due to expansion of time in other locks, a 128 partition run only spends 403 seconds less in locking than a 16 partition run. We can generate some OProfile statistics, but most of the time saved is probably absorbed into functions such as HeapTupleSatisfiesSnapshot and PinBuffer which seem to have a very high overhead. David -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of Simon Riggs Sent: Tuesday, September 12, 2006 1:37 AM To: Tom Lane Cc: Mark Wong; Bruce Momjian; PostgreSQL-development Subject: Re: [HACKERS] Lock partitions On Mon, 2006-09-11 at 11:29 -0400, Tom Lane wrote: Mark Wong [EMAIL PROTECTED] writes: Tom Lane wrote: It would be nice to see some results from the OSDL tests with, say, 4, 8, and 16 lock partitions before we forget about the point though. Anybody know whether OSDL is in a position to run tests for us? Yeah, I can run some dbt2 tests in the lab. I'll get started on it. We're still a little bit away from getting the automated testing for PostgreSQL going again though. Great, thanks. The thing to twiddle is LOG2_NUM_LOCK_PARTITIONS in src/include/storage/lwlock.h. You need a full backend recompile after changing it, but you shouldn't need to initdb, if that helps. IIRC we did that already and the answer was 16... -- Simon Riggs EnterpriseDB http://www.enterprisedb.com ---(end of broadcast)--- TIP 6: explain analyze is your friend ---(end of broadcast)--- TIP 1: if posting/reading through Usenet, please send an appropriate subscribe-nomail command to [EMAIL PROTECTED] so that your message can get through to the mailing list cleanly
Re: [HACKERS] Lock partitions
Strong, David [EMAIL PROTECTED] writes: When using 16 buffer and 16 lock partitions, we see that BufMapping takes 809 seconds to acquire locks and 174 seconds to release locks. The LockMgr takes 362 seconds to acquire locks and 26 seconds to release locks. When using 128 buffer and 128 lock partitions, we see that BufMapping takes 277 seconds (532 seconds improvement) to acquire locks and 78 seconds (96 seconds improvement) to release locks. The LockMgr takes 235 seconds (127 seconds improvement) to acquire locks and 22 seconds (4 seconds improvement) to release locks. While I don't see any particular penalty to increasing NUM_BUFFER_PARTITIONS, increasing NUM_LOCK_PARTITIONS carries a very significant penalty (increasing PGPROC size as well as the work needed during LockReleaseAll, which is executed at every transaction end). I think 128 lock partitions is probably verging on the ridiculous ... particularly if your benchmark only involves touching half a dozen tables. I'd be more interested in comparisons between 4 and 16 lock partitions. Also, please vary the two settings independently rather than confusing the issue by changing them both at once. With the improvements in the various locking times, one might expect an improvement in the overall benchmark result. However, a 16 partition run produces a result of 198.74 TPS and a 128 partition run produces a result of 203.24 TPS. Part of the time saved from BufMapping and LockMgr partitions is absorbed into the WALInsertLock lock. For a 16 partition run, the total time to lock/release the WALInsertLock lock is 5845 seconds. For 128 partitions, the WALInsertLock lock takes 6172 seconds, an increase of 327 seconds. Perhaps we have our WAL configured incorrectly? I fear this throws your entire measurement procedure into question. For a fixed workload the number of acquisitions of WALInsertLock ought to be fixed, so you shouldn't see any more contention for WALInsertLock if the transaction rate didn't change materially. regards, tom lane ---(end of broadcast)--- TIP 3: Have you checked our extensive FAQ? http://www.postgresql.org/docs/faq
Re: [HACKERS] Lock partitions
Tom, Thanks for the feedback. We'll run a few tests with differing buffer and lock partition sizes in the range you're interested in and we'll let you know what we see. Our workload is not fixed, however. Our benchmark does not follow the strict TPC-C guideline of using think times etc. We throw as many transactions at the database as we can. So, when any time is freed up, we will fill it with another transaction. We simply want to stress as much as we can. As one bottleneck is removed, the time saved obviously flows to the next. Postgres 8.2 moves some of the time that used to be consumed by single BufMappingLock and LockMGRLock locks to the WALInsertLock lock. We have run tests where we made XLogInsert a NOP, because we wanted to see where the next bottleneck would be, and some of the time occupied by WALInsertLock lock was absorbed by the SInvalLock lock. We have not tried to remove the SInvalLock lock to see where time flows to next, but we might. David -Original Message- From: Tom Lane [mailto:[EMAIL PROTECTED] Sent: Tuesday, September 12, 2006 9:40 AM To: Strong, David Cc: PostgreSQL-development Subject: Re: [HACKERS] Lock partitions Strong, David [EMAIL PROTECTED] writes: When using 16 buffer and 16 lock partitions, we see that BufMapping takes 809 seconds to acquire locks and 174 seconds to release locks. The LockMgr takes 362 seconds to acquire locks and 26 seconds to release locks. When using 128 buffer and 128 lock partitions, we see that BufMapping takes 277 seconds (532 seconds improvement) to acquire locks and 78 seconds (96 seconds improvement) to release locks. The LockMgr takes 235 seconds (127 seconds improvement) to acquire locks and 22 seconds (4 seconds improvement) to release locks. While I don't see any particular penalty to increasing NUM_BUFFER_PARTITIONS, increasing NUM_LOCK_PARTITIONS carries a very significant penalty (increasing PGPROC size as well as the work needed during LockReleaseAll, which is executed at every transaction end). I think 128 lock partitions is probably verging on the ridiculous ... particularly if your benchmark only involves touching half a dozen tables. I'd be more interested in comparisons between 4 and 16 lock partitions. Also, please vary the two settings independently rather than confusing the issue by changing them both at once. With the improvements in the various locking times, one might expect an improvement in the overall benchmark result. However, a 16 partition run produces a result of 198.74 TPS and a 128 partition run produces a result of 203.24 TPS. Part of the time saved from BufMapping and LockMgr partitions is absorbed into the WALInsertLock lock. For a 16 partition run, the total time to lock/release the WALInsertLock lock is 5845 seconds. For 128 partitions, the WALInsertLock lock takes 6172 seconds, an increase of 327 seconds. Perhaps we have our WAL configured incorrectly? I fear this throws your entire measurement procedure into question. For a fixed workload the number of acquisitions of WALInsertLock ought to be fixed, so you shouldn't see any more contention for WALInsertLock if the transaction rate didn't change materially. regards, tom lane ---(end of broadcast)--- TIP 6: explain analyze is your friend
Re: [HACKERS] Lock partitions
Mark Wong [EMAIL PROTECTED] writes: Tom Lane wrote: It would be nice to see some results from the OSDL tests with, say, 4, 8, and 16 lock partitions before we forget about the point though. Anybody know whether OSDL is in a position to run tests for us? Yeah, I can run some dbt2 tests in the lab. I'll get started on it. We're still a little bit away from getting the automated testing for PostgreSQL going again though. Great, thanks. The thing to twiddle is LOG2_NUM_LOCK_PARTITIONS in src/include/storage/lwlock.h. You need a full backend recompile after changing it, but you shouldn't need to initdb, if that helps. regards, tom lane ---(end of broadcast)--- TIP 2: Don't 'kill -9' the postmaster
Re: [HACKERS] Lock partitions
Tom Lane wrote: Bruce Momjian [EMAIL PROTECTED] writes: I see this in the CVS commits for 8.2. Did we determine the proper number of lock partitions? Should it be based on the number of buffers or concurrent sessions allowed? No. NUM_LOCK_PARTITIONS needs to be a compile-time constant for a number of reasons, and there is absolutely zero evidence to justify making any effort (and spending any cycles) on a variable value. It would be nice to see some results from the OSDL tests with, say, 4, 8, and 16 lock partitions before we forget about the point though. Anybody know whether OSDL is in a position to run tests for us? Yeah, I can run some dbt2 tests in the lab. I'll get started on it. We're still a little bit away from getting the automated testing for PostgreSQL going again though. Mark ---(end of broadcast)--- TIP 3: Have you checked our extensive FAQ? http://www.postgresql.org/docs/faq
Re: [HACKERS] Lock partitions
Bruce Momjian [EMAIL PROTECTED] writes: I see this in the CVS commits for 8.2. Did we determine the proper number of lock partitions? Should it be based on the number of buffers or concurrent sessions allowed? No. NUM_LOCK_PARTITIONS needs to be a compile-time constant for a number of reasons, and there is absolutely zero evidence to justify making any effort (and spending any cycles) on a variable value. It would be nice to see some results from the OSDL tests with, say, 4, 8, and 16 lock partitions before we forget about the point though. Anybody know whether OSDL is in a position to run tests for us? regards, tom lane ---(end of broadcast)--- TIP 6: explain analyze is your friend
