Thread: low performance
I am running the precomplied binary of Postgreql 7.1.2 on a Redhat 7.1 (on a Dual Celeron System with 256MB, kernel 2.4.4and 2.4.5) System. (The installation of the new 7.1.3 doesn't seem to solve the problem) I am connecting to the DB with a Perl Program (using Perl 5.6.0 with DBD-Pg-1.01 and DBI-1.19). The program inserts some million rows into a db with about 30 tables. The processing takes (if everyting works fine) about10 hours to complete. Usually the my Perl-Script and the database share the available CPU time 50:50. But sometimes the database is very slow eating up most (>98%) of the available CPU time. (Of cause I know VACUUM and VACUUM ANALYZE, this is not the problem). The only thing that seems to help then, is killing the perl script, stopping postgresql, running "ipcclean", and start againfrom the beginning. If it works from the beginning, the database is ususally very fast until all data are processed. But if someone else connects (using psql), sometimes the database gets very slow until it is using all the CPU time. There are no error messages at postgres-startup. I already increased the number of buffers to 2048 (doesn't help) I cannot reproduce these problems, sometimes the db is fast, sometimes very slow. The perl script doesn't seem to be theproblem, because I wrote all SQL Commands to a file and processed them later ("psql dbname postgres < SQL-File"). Same thing: sometimes slow sometimes fast. Andreas
I am aware of the performance drawbacks because of indices and triggers. In fact I have a trigger and an index on the mostpopulated table. It is not possible in my case to remove the primary keys during insert, because the database structure and foreign keys validatemy data during import. The problem is, that sometimes the performance is good, and sometimes the database is awfully slow. If it is slow, postgres is eating up all CPU time and it takes at least 150 times longer to insert the data. I don't know why and what to do against that. Andreas On Mon, 20 Aug 2001 19:39:31 -0400 Jonas Lindholm <jlindholm@rcn.com> wrote: > Do you have any index on the tables ? Any triggers ? > > If you want to insert 1 million rows you should drop the index, insert the data and then recreate the index. > You should also try the COPY command to insert the data. > > You should also avoid having anyone to connect to the database when you insert a lot of rows, and 1 million rows are alot of rows for any database. > > I've been able to insert, in one table, 17 million record in ~3 hours on a Compaq SMP 750 Mhz with 512MB > by dropping the index, using several COPY commands at the same time loading different parts of the data and then creatingthe index again. > At the time of the inserts no other processes than the COPY's was connected to the database. > > /Jonas Lindholm > > > Andreas Wernitznig wrote: > > > I am running the precomplied binary of Postgreql 7.1.2 on a Redhat 7.1 (on a Dual Celeron System with 256MB, kernel 2.4.4and 2.4.5) System. > > (The installation of the new 7.1.3 doesn't seem to solve the problem) > > > > I am connecting to the DB with a Perl Program (using Perl 5.6.0 with DBD-Pg-1.01 and DBI-1.19). > > The program inserts some million rows into a db with about 30 tables. The processing takes (if everyting works fine)about 10 hours to complete. Usually the my Perl-Script and the database share the available CPU time 50:50. > > But sometimes the database is very slow eating up most (>98%) of the available CPU time. > > (Of cause I know VACUUM and VACUUM ANALYZE, this is not the problem). > > > > The only thing that seems to help then, is killing the perl script, stopping postgresql, running "ipcclean", and startagain from the beginning. If it works from the beginning, the database is ususally very fast until all data are processed. > > > > But if someone else connects (using psql), sometimes the database gets very slow until it is using all the CPU time. > > > > There are no error messages at postgres-startup. > > I already increased the number of buffers to 2048 (doesn't help) > > > > I cannot reproduce these problems, sometimes the db is fast, sometimes very slow. The perl script doesn't seem to bethe problem, because I wrote all SQL Commands to a file and processed them later ("psql dbname postgres < SQL-File"). > > Same thing: sometimes slow sometimes fast. > > > > Andreas >
Andreas Wernitznig <andreas@insilico.com> writes: > I am aware of the performance drawbacks because of indices and > triggers. In fact I have a trigger and an index on the most populated > table. It is not possible in my case to remove the primary keys > during insert, because the database structure and foreign keys > validate my data during import. Foreign keys eh? > The problem is, that sometimes the performance is good, and sometimes > the database is awfully slow. If it is slow, postgres is eating up > all CPU time and it takes at least 150 times longer to insert the > data. I don't know why and what to do against that. We found some foreign-key-related performance problems not long ago, and it could be you're happening on another one. However there's not enough info here to figure it out. I can offer you two alternatives: 1. Compile up the backend with profiling enabled (if you're using gcc then "make PROFILE=-pg clean all" in src/backend should do the trick). Collect profiles for both a "normal" and a "slow" run and send them in. 2. Develop a self-contained example that exhibits the problem, and send it along for someone else to profile. regards, tom lane
I took option 1 and managed to create a profile of a slow and a fast run: The frequent functions of the FAST run: % cumulative self self total time seconds seconds calls Ts/call Ts/call name 0.00 0.00 0.00 15725437 0.00 0.00 AllocSetAlloc 0.00 0.00 0.00 15380742 0.00 0.00 MemoryContextAlloc 0.00 0.00 0.00 11296700 0.00 0.00 ExecEvalExpr 0.00 0.00 0.00 8276639 0.00 0.00 newNode 0.00 0.00 0.00 5430717 0.00 0.00 MemoryContextSwitchTo 0.00 0.00 0.00 4492641 0.00 0.00 LockBuffer 0.00 0.00 0.00 4425642 0.00 0.00 AllocSetFree 0.00 0.00 0.00 4356571 0.00 0.00 pfree 0.00 0.00 0.00 3873174 0.00 0.00 pq_getbyte 0.00 0.00 0.00 3799725 0.00 0.00 appendStringInfoChar The frequent functions of the SLOW run: % cumulative self self total time seconds seconds calls Ts/call Ts/call name 0.00 0.00 0.00 27832819 0.00 0.00 ExecEvalExpr 0.00 0.00 0.00 19040887 0.00 0.00 AllocSetAlloc 0.00 0.00 0.00 18976313 0.00 0.00 MemoryContextAlloc 0.00 0.00 0.00 18722462 0.00 0.00 LockBuffer 0.00 0.00 0.00 18684832 0.00 0.00 MemoryContextSwitchTo 0.00 0.00 0.00 18442039 0.00 0.00 pg_detoast_datum 0.00 0.00 0.00 16947638 0.00 0.00 AllocSetFree 0.00 0.00 0.00 16934648 0.00 0.00 pfree 0.00 0.00 0.00 9716164 0.00 0.00 SpinAcquire 0.00 0.00 0.00 9716164 0.00 0.00 SpinRelease Since these files are to big for a posting, I have put the whole profile files on: ftp://ftp.insilico.com/out.fast.gz ftp://ftp.insilico.com/out.slow.gz I don't know why the time column and number of seconds is zero in all the cases. I am using the Redhat 7.1 binutils (binutils-2.10.91.0.2-3). On Tue, 21 Aug 2001 17:38:23 -0400 Tom Lane <tgl@sss.pgh.pa.us> wrote: > Andreas Wernitznig <andreas@insilico.com> writes: > > I am aware of the performance drawbacks because of indices and > > triggers. In fact I have a trigger and an index on the most populated > > table. It is not possible in my case to remove the primary keys > > during insert, because the database structure and foreign keys > > validate my data during import. > > Foreign keys eh? > > > The problem is, that sometimes the performance is good, and sometimes > > the database is awfully slow. If it is slow, postgres is eating up > > all CPU time and it takes at least 150 times longer to insert the > > data. I don't know why and what to do against that. > > We found some foreign-key-related performance problems not long ago, > and it could be you're happening on another one. However there's not > enough info here to figure it out. I can offer you two alternatives: > > 1. Compile up the backend with profiling enabled (if you're using gcc > then "make PROFILE=-pg clean all" in src/backend should do the trick). > Collect profiles for both a "normal" and a "slow" run and send them in. > > 2. Develop a self-contained example that exhibits the problem, and send > it along for someone else to profile. > > regards, tom lane > > ---------------------------(end of broadcast)--------------------------- > TIP 5: Have you checked our extensive FAQ? > > http://www.postgresql.org/users-lounge/docs/faq.html >
Andreas Wernitznig <andreas@insilico.com> writes: > I took option 1 and managed to create a profile of a slow and a fast run: It's difficult to compare these profiles, because they seem to be taken over very different numbers of queries --- did you let the "fast" run process more queries than the "slow" one? However, I think what is happening is that some queries are being done as indexscans in the fast case and seqscans in the slow case. The ratio of ExecIndexScan calls to ExecSeqScan calls is vastly different in the two profiles. It looks like you are doing vacuums fairly frequently, so I speculate that the statistics gathered by vacuum are changing just enough to alter the planner's choice between indexscan and seqscan for some often-executed query type. Evidently the planner is guessing the costs to be nearly the same, so a small change in stats might tip the choice --- but in reality the costs are quite a bit different, thus you observe fast and slow behavior. The next step should be to get EXPLAIN results for the queries used by your application in both fast and slow states. This will help us narrow down where the planner's misprediction is occurring. regards, tom lane
Andreas Wernitznig <andreas@insilico.com> writes: > To make it more comparable I have made two additional runs, a slow and > a fast one with exactly the same number of inserts (about 20500) and > put it on our ftp server: >> However, I think what is happening is that some queries are being done >> as indexscans in the fast case and seqscans in the slow case. The >> ratio of ExecIndexScan calls to ExecSeqScan calls is vastly different >> in the two profiles. > Does the new profiles proof that assumption ? Yes, see for yourself: def.fast: 0.00 0.00 0.00 22481 0.00 0.00 ExecSeqScan 0.00 0.00 0.00 20161 0.00 0.00 ExecIndexScan def.slow: 0.00 0.01 0.00 41940 0.00 0.00 ExecSeqScan 0.00 0.01 0.00 702 0.00 0.00 ExecIndexScan So there are about 19500 queries that are being done as indexscans in one case and seqscans in the other. > If I run "vacuum" and "vacuum analyze" on an empty database, the > following run will be a SLOW one. The whole point of vacuum analyze is to give the planner some statistics about the contents of the tables. Vacuum analyze when a table is empty is useless (even counterproductive, if the table shortly thereafter becomes large --- the planner will still think it is empty). regards, tom lane
Andreas Wernitznig <andreas@insilico.com> writes: > The only way to make it faster after step 3 is to close that connection (and stop that postmaster thread with it) and establisha new one. > It seems like the planner (at least for pk checking) of an *established* connection to a database doesn't receive the informationgained from "vacuum analyze". Ah. Yes, you are quite right: the queries used for PK checking are planned just once when first executed, and thereafter the plans are cached and reused for the life of that backend. No doubt you are seeing continued use of a no-longer-appropriate PK plan. We have a TODO item to be smarter about dropping cached plans when relevant context changes, but it's not done yet. regards, tom lane
Yes, I understand very clearly what you mean. Maybe my mails were to confused, that's why I try to explain my problem once more: step 1. An empty table with a primary key (=index key) where an "explain" tells me, that a Seq Scan is used to SELECT a specialrow. step 2. Then I start to fill data into that table. step 3. Then I run a vacuum analyze to update the planner statistics. step 4. I run an "EXPLAIN select * from <mytable> where <pk-column> = 999;" step 5. Then I fill in additional data. What I expect is, that from step 5 on the pk-trigger (I don't know if this mechanism that checks for uniqueness is reallya trigger) uses the Index to check for possible double entries. Although "EXPLAIN" in step 4 pretend to use an Index Scan the data insert becomes slower and slower (>98% of the processoris used by a postmaster). All these steps are done with a single connection (postmaster). The only way to make it faster after step 3 is to close that connection (and stop that postmaster thread with it) and establisha new one. It seems like the planner (at least for pk checking) of an *established* connection to a database doesn't receive the informationgained from "vacuum analyze". Greetings Andreas On Mon, 03 Sep 2001 12:26:39 -0400 Tom Lane <tgl@sss.pgh.pa.us> wrote: > Andreas Wernitznig <andreas@insilico.com> writes: > > To make it more comparable I have made two additional runs, a slow and > > a fast one with exactly the same number of inserts (about 20500) and > > put it on our ftp server: > > >> However, I think what is happening is that some queries are being done > >> as indexscans in the fast case and seqscans in the slow case. The > >> ratio of ExecIndexScan calls to ExecSeqScan calls is vastly different > >> in the two profiles. > > > Does the new profiles proof that assumption ? > > Yes, see for yourself: > def.fast: > 0.00 0.00 0.00 22481 0.00 0.00 ExecSeqScan > 0.00 0.00 0.00 20161 0.00 0.00 ExecIndexScan > def.slow: > 0.00 0.01 0.00 41940 0.00 0.00 ExecSeqScan > 0.00 0.01 0.00 702 0.00 0.00 ExecIndexScan > > So there are about 19500 queries that are being done as indexscans in > one case and seqscans in the other. > > > If I run "vacuum" and "vacuum analyze" on an empty database, the > > following run will be a SLOW one. > > The whole point of vacuum analyze is to give the planner some statistics > about the contents of the tables. Vacuum analyze when a table is empty > is useless (even counterproductive, if the table shortly thereafter > becomes large --- the planner will still think it is empty). > > regards, tom lane >