Thread: how much ram do i give postgres?
I know this is kinda a debate, but how much ram do I give postgres? I've seen many places say around 10-15% or some say 25%....... If all this server is doing is running postgres, why can't I give it 75%+? Should the limit be as much as possible as long as the server doesn't use any swap? Any thoughts would be great, but I'd like to know why. Thanks. -Josh
Josh Close <narshe@gmail.com> writes:
> I know this is kinda a debate, but how much ram do I give postgres?
> I've seen many places say around 10-15% or some say 25%....... If all
> this server is doing is running postgres, why can't I give it 75%+?
> Should the limit be as much as possible as long as the server doesn't
> use any swap?
The short answer is no; the sweet spot for shared_buffers is usually on
the order of 10000 buffers, and trying to go for "75% of RAM" isn't
going to do anything except hurt. For the long answer see the
pgsql-performance list archives.
regards, tom lane
On Tue, 19 Oct 2004 17:42:16 -0400, Tom Lane <tgl@sss.pgh.pa.us> wrote: > The short answer is no; the sweet spot for shared_buffers is usually on > the order of 10000 buffers, and trying to go for "75% of RAM" isn't > going to do anything except hurt. For the long answer see the > pgsql-performance list archives. > > regards, tom lane Well, I didn't find a whole lot in the list-archives, so I emailed that list whith a few more questions. My postgres server is just crawling right now :( -Josh
On 19 Oct 2004 at 17:35, Josh Close wrote: > Well, I didn't find a whole lot in the list-archives, so I emailed > that list whith a few more questions. My postgres server is just > crawling right now :( > Unlike many other database engines the shared buffers of Postgres is not a private cache of the database data. It is a working area shared between all the backend processes. This needs to be tuned for number of connections and overall workload, *not* the amount of your database that you want to keep in memory. There is still lots of debate about what the "sweet spot" is. Maybe there isn't one, but its not normally 75% of RAM. If anything, the effective_cache_size needs to be 75% of (available) RAM as this is telling Postgres the amount of your database the *OS* is likely to cache in memory. Having said that, I think you will need to define "crawling". Is it updates/inserts that are slow? This may be triggers/rules/referential integrity checking etc that is slowing it. If it is selects that are slow, this may be incorrect indexes or sub-optimal queries. You need to show us what you are trying to do and what the results are. Regards, Gary.
On Wed, 20 Oct 2004 08:00:55 +0100, Gary Doades <gpd@gpdnet.co.uk> wrote: > Unlike many other database engines the shared buffers of Postgres is > not a private cache of the database data. It is a working area shared > between all the backend processes. This needs to be tuned for number > of connections and overall workload, *not* the amount of your database > that you want to keep in memory. There is still lots of debate about what > the "sweet spot" is. Maybe there isn't one, but its not normally 75% of > RAM. > > If anything, the effective_cache_size needs to be 75% of (available) > RAM as this is telling Postgres the amount of your database the *OS* is > likely to cache in memory. > > Having said that, I think you will need to define "crawling". Is it > updates/inserts that are slow? This may be triggers/rules/referential > integrity checking etc that is slowing it. If it is selects that are slow, this > may be incorrect indexes or sub-optimal queries. You need to show us > what you are trying to do and what the results are. It's slow due to several things happening all at once. There are a lot of inserts and updates happening. There is periodically a bulk insert of 500k - 1 mill rows happening. I'm doing a vacuum anaylyze every hour due to the amount of transactions happening, and a vacuum full every night. All this has caused selects to be very slow. At times, a "select count(1)" from a table will take several mins. I don't think selects would have to wait on locks by inserts/updates would it? I would just like to do anything possible to help speed this up. -Josh
>It's slow due to several things happening all at once. There are a lot >of inserts and updates happening. There is periodically a bulk insert >of 500k - 1 mill rows happening. I'm doing a vacuum anaylyze every >hour due to the amount of transactions happening, and a vacuum full >every night. All this has caused selects to be very slow. At times, a >"select count(1)" from a table will take several mins. I don't think >selects would have to wait on locks by inserts/updates would it? > >I would just like to do anything possible to help speed this up. > > If there are really many rows in table , select count(1) would be a little bit slow, for postgresql use sequential scan to count the rows. If the query is other kind, then may be check if there are index on search condition or use EXPLAIN command to see the query plan would be greatly help. By the way, what's the version of your postgresql? older version (<7.4?) still suffer from index space bloating. regards Laser
On Wed, 2004-10-20 at 07:25, Josh Close wrote: > It's slow due to several things happening all at once. There are a lot > of inserts and updates happening. There is periodically a bulk insert > of 500k - 1 mill rows happening. I'm doing a vacuum anaylyze every > hour due to the amount of transactions happening, and a vacuum full > every night. All this has caused selects to be very slow. At times, a > "select count(1)" from a table will take several mins. I don't think > selects would have to wait on locks by inserts/updates would it? 1: Is the bulk insert being done inside of a single transaction, or as individual inserts? 2: Are your fsm settings high enough for an hourly vacuum to be effective? 3: How selective is the where clause for your select (1) query? If there is no where clause or the where clause isn't very selective, then there will be a sequential scan every time. Since PostgreSQL has to hit the table after using an index anyway, if it's going to retrieve a fair percent of a table, it just goes right to a seq scan, which for postgresql, is the right thing to do. Post "explain analyze" of your slowest queries to the performance list if you can.
On Wed, 20 Oct 2004 09:52:25 -0600, Scott Marlowe <smarlowe@qwest.net> wrote: > 1: Is the bulk insert being done inside of a single transaction, or as > individual inserts? The bulk insert is being done by COPY FROM STDIN. It copies in 100,000 rows at a time, then disconnects, reconnects, and copies 100k more, and repeats 'till done. There are no indexes on the tables that the copy is being done into either, so it won't be slowed down by that at all. > > 2: Are your fsm settings high enough for an hourly vacuum to be > effective? What is fsm? I'll tell you when I find that out. > > 3: How selective is the where clause for your select (1) query? If > there is no where clause or the where clause isn't very selective, then > there will be a sequential scan every time. Since PostgreSQL has to hit > the table after using an index anyway, if it's going to retrieve a fair > percent of a table, it just goes right to a seq scan, which for > postgresql, is the right thing to do. There was no where clause. > > Post "explain analyze" of your slowest queries to the performance list > if you can. I don't think it's a query problem ( but I could optimize them more I'm sure ), 'cause the same query takes a long time when there are other queries happening, and not long at all when nothing else is going on. -Josh
On 20 Oct 2004 at 11:37, Josh Close wrote: > On Wed, 20 Oct 2004 09:52:25 -0600, Scott Marlowe <smarlowe@qwest.net> wrote: > > 1: Is the bulk insert being done inside of a single transaction, or as > > individual inserts? > > The bulk insert is being done by COPY FROM STDIN. It copies in 100,000 > rows at a time, then disconnects, reconnects, and copies 100k more, > and repeats 'till done. There are no indexes on the tables that the > copy is being done into either, so it won't be slowed down by that at > all. > > > What about triggers? Also constraints (check contraints, integrity constraints) All these will slow the inserts/updates down. If you have integrity constraints make sure you have indexes on the referenced columns in the referenced tables and make sure the data types are the same. How long does 100,000 rows take to insert exactly? How many updates are you performing each hour? Regards, Gary.
On Wed, Oct 20, 2004 at 08:25:22 -0500, Josh Close <narshe@gmail.com> wrote: > > It's slow due to several things happening all at once. There are a lot > of inserts and updates happening. There is periodically a bulk insert > of 500k - 1 mill rows happening. I'm doing a vacuum anaylyze every > hour due to the amount of transactions happening, and a vacuum full > every night. All this has caused selects to be very slow. At times, a > "select count(1)" from a table will take several mins. I don't think > selects would have to wait on locks by inserts/updates would it? You might not need to do the vacuum fulls that often. If the your hourly vacuums have a high enough fsm setting, they should be keeping the database from continually growing in size. At that point daily vacuum fulls are overkill and if they are slowing stuff down you want to run quickly, you should cut back on them.
On Wed, 20 Oct 2004 18:47:25 +0100, Gary Doades <gpd@gpdnet.co.uk> wrote: > What about triggers? Also constraints (check contraints, integrity > constraints) All these will slow the inserts/updates down. No triggers or constraints. There are some foreign keys, but the tables that have the inserts don't have anything to them, even indexes, to help speed up the inserts. > > If you have integrity constraints make sure you have indexes on the > referenced columns in the referenced tables and make sure the data > types are the same. > > How long does 100,000 rows take to insert exactly? I believe with the bulk inserts, 100k only takes a couple mins. > > How many updates are you performing each hour? I'm not sure about this. Is there a pg stats table I can look at to find this out..... I suppose I could do a count on the time stamp also. I'll let you know when I find out. > > Regards, > Gary. > > ---------------------------(end of broadcast)--------------------------- > TIP 4: Don't 'kill -9' the postmaster >
On Wed, 20 Oct 2004 13:35:43 -0500, Bruno Wolff III <bruno@wolff.to> wrote: > You might not need to do the vacuum fulls that often. If the your hourly > vacuums have a high enough fsm setting, they should be keeping the database > from continually growing in size. At that point daily vacuum fulls are > overkill and if they are slowing stuff down you want to run quickly, you > should cut back on them. I have the vacuum_mem set at 32M right now. I haven't changed the fsm settings at all though. -Josh
On 20 Oct 2004 at 13:34, Josh Close wrote: > > How long does 100,000 rows take to insert exactly? > > I believe with the bulk inserts, 100k only takes a couple mins. > Hmm, that seems a bit slow. How big are the rows you are inserting? Have you checked the cpu and IO usage during the inserts? You will need to do some kind of cpu/IO monitoring to determine where the bottleneck is. What hardware is this on? Sorry if you specified it earlier, I can't seem to find mention of it. Cheers, Gary.
On Wed, 20 Oct 2004 19:59:38 +0100, Gary Doades <gpd@gpdnet.co.uk> wrote: > Hmm, that seems a bit slow. How big are the rows you are inserting? Have you checked > the cpu and IO usage during the inserts? You will need to do some kind of cpu/IO > monitoring to determine where the bottleneck is. The bulk inserts don't take full cpu. Between 40% and 80%. On the other hand, a select will take 99% cpu. > > What hardware is this on? Sorry if you specified it earlier, I can't seem to find mention of > it. It's on a P4 HT with 1,128 megs ram. -Josh
On 20 Oct 2004 at 14:09, Josh Close wrote: > On Wed, 20 Oct 2004 19:59:38 +0100, Gary Doades <gpd@gpdnet.co.uk> wrote: > > Hmm, that seems a bit slow. How big are the rows you are inserting? Have you checked > > the cpu and IO usage during the inserts? You will need to do some kind of cpu/IO > > monitoring to determine where the bottleneck is. > > The bulk inserts don't take full cpu. Between 40% and 80%. On the > other hand, a select will take 99% cpu. Is this the select(1) query? Please post an explain analyze for this and any other "slow" queries. I would expect the selects to take 99% cpu if all the data you were trying to select was already in memory. Is this the case in general? I can do a "select count(1)" on a 500,000 row table in about 1 second on a Athlon 2800+ if all the data is cached. It takes about 25 seconds if it has to fetch it from disk. I have just done a test by inserting (via COPY) of 149,000 rows in a table with 23 columns, mostly numeric, some int4, 4 timestamps. This took 28 seconds on my Windows XP desktop, Athlon 2800+, 7200 rpm SATA disk, Postgres 8.0 beta 2. It used around 20% to 40% cpu during the copy. The only index was the int4 primary key, nothing else. How does this compare? > > What hardware is this on? Sorry if you specified it earlier, I can't seem to find mention of > > it. > > It's on a P4 HT with 1,128 megs ram. Disk system?? Regards, Gary.
On Wed, 20 Oct 2004 20:49:54 +0100, Gary Doades <gpd@gpdnet.co.uk> wrote: > Is this the select(1) query? Please post an explain analyze for this and any other "slow" > queries. I think it took so long 'cause it wasn't cached. The second time I ran it, it took less than a second. How you can tell if something is cached? Is there a way to see what's in cache? > > I would expect the selects to take 99% cpu if all the data you were trying to select was > already in memory. Is this the case in general? I can do a "select count(1)" on a 500,000 > row table in about 1 second on a Athlon 2800+ if all the data is cached. It takes about 25 > seconds if it has to fetch it from disk. I think that's what's going on here. > > I have just done a test by inserting (via COPY) of 149,000 rows in a table with 23 > columns, mostly numeric, some int4, 4 timestamps. This took 28 seconds on my > Windows XP desktop, Athlon 2800+, 7200 rpm SATA disk, Postgres 8.0 beta 2. It used > around 20% to 40% cpu during the copy. The only index was the int4 primary key, > nothing else. Well, there are a 3 text columns or so, and that's why the COPY takes longer than yours. That hasn't been a big issue though. I copies fast enough. > > How does this compare? > > Disk system?? It's in ide raid 1 config I believe. So it's not too fast. It will soon be on a scsi raid 5 array. That should help speed some things up also. > > Regards, > Gary. What about the postgresql.conf config settings. This is what I have and why. shared_buffers = 21250 This is 174 megs, which is 15% of total ram. I read somewhere that it should be between 12-15% of total ram. sort_mem = 32768 This is default. vacuum_mem = 32768 This is 32 megs. I put it that high because of something I read here http://www.varlena.com/varlena/GeneralBits/Tidbits/perf.html #max_fsm_pages = 20000 Default. I would think this could be upped more, but I don't know how much. effective_cache_size = 105750 This is 846 megs ram which is 75% of total mem. I put it there 'cause of a reply I got on the performance list. I made all these changes today, and haven't had much of a chance to speed test postgres since. Any thoughs on these settings? -Josh
On 20 Oct 2004 at 15:36, Josh Close wrote: > On Wed, 20 Oct 2004 20:49:54 +0100, Gary Doades <gpd@gpdnet.co.uk> wrote: > > Is this the select(1) query? Please post an explain analyze for this and any other "slow" > > queries. > > I think it took so long 'cause it wasn't cached. The second time I ran > it, it took less than a second. How you can tell if something is > cached? Is there a way to see what's in cache? No. The OS caches the data as read from the disk. If you need the data to be in memory for performance then you need to make sure you have enough available RAM to hold your typical result sets if possible. > What about the postgresql.conf config settings. This is what I have and why. > > sort_mem = 32768 > > This is default. This is not the default. The default is 1000. You are telling Postgres to use 32Megs for *each* sort that is taking place. If you have several queries each performing large sorts you can quickly eat up available RAM this way. If you will only have a small number of concurrrent queries performing sorts then this may be OK. Don't forget, a single query can perform more than one sort operation. If you have 10 large sorts happening at the same time, you can eat up to 320 megs this way! You will need to tell us the number of updates/deletes you are having. This will determine the vacuum needs. If the bulk of the data is inserted you may only need to analyze frequently, not vacuum. In order to get more help you will need to supply the update/delete frequency and the explain analyze output from your queries. Regards, Gary.
On Wed, 20 Oct 2004 23:43:54 +0100, Gary Doades <gpd@gpdnet.co.uk> wrote: > You will need to tell us the number of updates/deletes you are having. This will > determine the vacuum needs. If the bulk of the data is inserted you may only need to > analyze frequently, not vacuum. > > In order to get more help you will need to supply the update/delete frequency and the > explain analyze output from your queries. I will have to gather this information for you. -Josh