Обсуждение: Performance considerations for very heavy INSERT traffic

I'm in the process of developing an application which uses PostgreSQL for data storage.  Our database traffic is very atypical, and as a result it has been rather challenging to figure out how to best tune PostgreSQL on what development hardware we have, as well as to figure out exactly what we should be evaluating and eventually buying for production hardware.

The vast, overwhelming majority of our database traffic is pretty much a non-stop stream of INSERTs filling up tables.  It is akin to data acquisition.  Several thousand clients are sending once-per-minute updates full of timestamped numerical data at our central server, which in turn performs INSERTs into several distinct tables as part of the transaction for that client.  We're talking on the order of ~100 transactions per second, each containing INSERTs to multiple tables (which contain only integer and floating point columns and a timestamp column - the primary key (and only index) is on a unique integer ID for the client and the timestamp).  The transaction load is spread evenly over time by having the clients send their per-minute updates at random times rather than on the exact minute mark.

There will of course be users using a web-based GUI to extract data from these tables and display them in graphs and whatnot, but the SELECT query traffic will always be considerably less frequent and intensive than the incessant INSERTs, and it's not that big a deal if the large queries take a little while to run.

This data also expires - rows with timestamps older than X days will be DELETEd periodically (once an hour or faster), such that the tables will reach a relatively stable size (pg_autovacuum is handling vacuuming for now, but considering our case, we're thinking of killing pg_autovacuum in favor of having the periodic DELETE process also do a vacuum of affected tables right after the DELETE, and then have it vacuum the other low traffic tables once a day while it's at it).

There is an aggregation layer in place which proxies the inbound data from the clients into a small(er) number of persistent postgresql backend processes.  Right now we're doing one aggregator per 128 clients (so instead of 128 seperate database connections over the course of a minute for a small transaction each, there is a single database backend that is constantly committing transactions at a rate of ~ 2/second).  At a test load of ~1,000 clients, we would have 8 aggregators running and 8 postgresql backends.  Testing has seemed to indicate we should aggregate even harder - the planned production load is ~5,000 clients initially, but will grow to almost double that in the not-too-distant future, and that would mean ~40 backends at 128 clients each initially.  Even on 8 cpus, I'm betting 40 concurrent backends doing 2 tps is much worse off than 10 backends doing 8 tps.

Test hardware right now is a dual Opteron with 4G of ram, which we've barely gotten 1,000 clients running against.  Current disk hardware in testing is whatever we could scrape together (4x 3-ware PCI hardware RAID controllers, with 8 SATA drives in a RAID10 array off of each - aggregated up in a 4-way stripe with linux md driver and then formatted as ext3 with an appropriate stride parameter and data=writeback).  Production will hopefully be a 4-8-way Opteron, 16 or more G of RAM, and a fiberchannel hardware raid array or two (~ 1TB available RAID10 storage) with 15krpm disks and battery-backed write cache.

I know I haven't provided a whole lot of application-level detail here, but does anyone have any general advice on tweaking postgresql to deal with a very heavy load of concurrent and almost exclusively write-only transactions?  Increasing shared_buffers seems to always help, even out to half of the dev box's ram (2G).  A 100ms commit_delay seemed to help, but tuning it (and _siblings) has been difficult.  We're using 8.0 with the default 8k blocksize, but are strongly considering both developing against 8.1 (seems it might handle the heavy concurrency better), and re-compiling with 32k blocksize since our storage arrays will inevitably be using fairly wide stripes.  Any advice on any of this (other than drop the project while you're still a little bit sane)?

--Brandon

> I know I haven't provided a whole lot of application-level detail here,

    You did !

    What about :

    - using COPY instead of INSERT ?
        (should be easy to do from the aggregators)

    - using Bizgres ?
        (which was designed for your use case)

    - splitting the xlog and the data on distinct physical drives or arrays

    - benchmarking something else than ext3
        (xfs ? reiser3 ?)

"Brandon Black" <blblack@gmail.com> wrote ...

Increasing shared_buffers seems to always help, even out to half of the dev box's ram (2G). 

 

Though officially PG does not prefer huge shared_buffers size, I did see several times that performance was boosted in case IO is the bottleneck. Also, if you want to use big bufferpool setting, make sure your version has Tom's split BufMgrLock patch

(http://archives.postgresql.org/pgsql-committers/2005-03/msg00025.php), which might already in 8.0.x somewhere. And if you want to use bufferpool bigger than 2G on 64-bit machine, you may need 8.1 (http://archives.postgresql.org/pgsql-committers/2005-08/msg00221.php).

 

Regards,

Qingqing

Split your system into multiple partitions of RAID 10s.For max performance,  ten drive RAID 10 for pg_xlog (This will max out a PCI-X bus) on Bus A, multiple 4/6Drive RAID 10s for tablespaces on Bus B. For max performance I would recommend using one RAID 10 for raw data tables, one for aggregate tables and one for indexes.  More RAM will only help you with queries against your data, if you are pre-aggregating, then you may not need all that much RAM.

You can easily get 100 tansacts per second with even less hardware with a little data partitioning.

Choose your controller carefully as many don't co-operate with linux well.

Alex Turner
NetEconomist

We've had bad experiences under extreme and/or strange workloads with XFS here in general, although this is the first major postgresql project - the rest were with other applications writing to XFS.  Bad experiences like XFS filesystems "detecting internal inconsistencies" at runtime and unmounting themselves from within the kernel module (much to the dismay of applications with open files on the filesystem), on machines with validated good hardware.  It has made me leary of using anything other than ext3 for fear of stability problems.  Reiser3 might be worth taking a look at though.

Thanks for the ideas,
-- Brandon

Brandon Black wrote:

>
>
> On 9/12/05, *PFC* <lists@boutiquenumerique.com
> <mailto:lists@boutiquenumerique.com>> wrote:
>
>
>
>             - benchmarking something else than ext3
>                     (xfs ? reiser3 ?)
>
>
> We've had bad experiences under extreme and/or strange workloads with
> XFS here in general, although this is the first major postgresql
> project - the rest were with other applications writing to XFS.  Bad
> experiences like XFS filesystems "detecting internal inconsistencies"
> at runtime and unmounting themselves from within the kernel module
> (much to the dismay of applications with open files on the
> filesystem), on machines with validated good hardware.  It has made me
> leary of using anything other than ext3 for fear of stability
> problems.  Reiser3 might be worth taking a look at though.

Just one tidbit.   We tried XFS on a very active system similar to what
you describe.   Dual opterons, 8GB memory, fiber channel drives, 2.6
kernel, etc.   And the reliability was awful.   We spent a lot of time
making changes one at a time to try and isolate the cause; when we
switched out from XFS to ReiserFS our stability problems went away.

It may be the case that the XFS problems have all been corrected in
newer kernels, but I'm not going to put too much effort into trying that
again.


I recently built a postgres with 32KB block sizes and have been doing
some testing.  For our particular workloads it has been a win.

-- Alan

Brandon Black <blblack@gmail.com> writes:

> The vast, overwhelming majority of our database traffic is pretty much a
> non-stop stream of INSERTs filling up tables.

That part Postgres should handle pretty well. It should be pretty much limited
by your I/O bandwidth so big raid 1+0 arrays are ideal. Putting the WAL on a
dedicated array would also be critical.

The WAL parameters like commit_delay and commit_siblings are a bit of a
mystery. Nobody has done any extensive testing of them. It would be quite
helpful if you find anything conclusive and post it. It would also be
surprising if they had a very large effect. They almost got chopped recently
because they weren't believed to be useful.

You might also ponder whether you need to by issuing a commit for every datum.
If you only need to commit periodically you can get much better throughput. I
suppose that's the same as commit_siblings. It would be interesting to know if
you can get those parameters to perform as well as batching up records
yourself.

> There will of course be users using a web-based GUI to extract data from
> these tables and display them in graphs and whatnot, but the SELECT query
> traffic will always be considerably less frequent and intensive than the
> incessant INSERTs, and it's not that big a deal if the large queries take a
> little while to run.

I do fear these queries. Even if they aren't mostly terribly intensive if
you're pushing the edges of your write I/O bandwidth then a single seek to
satisfy one of these selects could really hurt your throughput.

That said, as long as your WAL is on a dedicated drive Postgres's architecture
should in theory be ideal and allow you do run these things with impunity. The
WAL is purely write-only and it's the only thing your inserts will be blocking
on.

> This data also expires - rows with timestamps older than X days will be
> DELETEd periodically (once an hour or faster), such that the tables will
> reach a relatively stable size (pg_autovacuum is handling vacuuming for now,
> but considering our case, we're thinking of killing pg_autovacuum in favor
> of having the periodic DELETE process also do a vacuum of affected tables
> right after the DELETE, and then have it vacuum the other low traffic tables
> once a day while it's at it).

Ay, there's the rub.

Taking this approach means you have vacuums running which have to scan your
entire table and your inserts are being sprayed all over the disk.

An alternative you may consider is using partitioned tables. Then when you
want to expire old records you simply drop the oldest partition. Or in your
case you might rotate through the partitions, so you just truncate the oldest
one and start inserting into it.

Unfortunately there's no built-in support for partitioned tables in Postgres.
You get to roll your own using UNION ALL or using inherited tables. Various
people have described success with each option though they both have
downsides too.

Using partitioned tables you would never need to delete any records except for
when you delete all of them. So you would never need to run vacuum except on
newly empty partitions. That avoids having to scan through all those pages
that you know will never have holes. If you use TRUNCATE (or VACUUM ALL or
CLUSTER) that would mean your inserts are always sequential too (though it
also means lots of block allocations along the way, possibly not an
advantage).

This may be a lot more work to set up and maintain but I think it would be a
big performance win. It would directly speed up the WAL writes by avoiding
those big full page dumps. And it would also cut out all the I/O traffic
vacuum generates.


> Increasing shared_buffers seems to always help, even out to half of the dev
> box's ram (2G).

Half should be a pessimal setting. It means virtually everything is buffered
twice. Once in the kernel and once in Postgres. Effectively halving your
memory. If that's really helping try raising it even further, to something
like 90% of your memory. But the conventional dogma is that shared buffers
should never be more than about 10k no matter how much memory you have. Let
the kernel do the bulk of the buffering.

That said it's even more mysterious in your situation. Why would a large
shared buffers help insert-only performance much at all? My guess is that it's
letting the vacuums complete quicker. Perhaps try raising work_mem?

--
greg

On 9/12/05, Brandon Black <blblack@gmail.com> wrote:
>
>  I'm in the process of developing an application which uses PostgreSQL for
> data storage.  Our database traffic is very atypical, and as a result it has
> been rather challenging to figure out how to best tune PostgreSQL on what
> development hardware we have, as well as to figure out exactly what we
> should be evaluating and eventually buying for production hardware.

A few suggestions...

1) Switch to COPY if you can, it's anywhere from 10-100x faster than
INSERT, but it does not necessarily fit your idea of updating multiple
tables.  In that case, try and enlarge the transaction's scope and do
multiple INSERTs in the same transaction.  Perhaps batching once per
second, or 5 seconds, and returning the aggregate result ot the
clients.

2) Tune ext3.  The default configuration wrecks high-write situations.
 Look into data=writeback for mounting, turning off atime (I hope
you've done this already) updates, and also modifying the scheduler to
the elevator model.  This is poorly documented in Linux (like just
about everything), but it's crtical.

3) Use 8.1 and strongly look at Bizgres. The data partitioning is critical.

4) Make sure you are not touching more data than you need, and don't
have any extraneous indexes.  Use the planner to make sure every index
is used, as it substantially increases the write load.

I've worked on a few similar applications, and this is a hard thing in
any database, even Oracle.

Chris

--
| Christopher Petrilli
| petrilli@gmail.com

* Brandon Black (blblack@gmail.com) wrote:
> Ideally I'd like to commit the data seperately, as the data could contain
> errors which abort the transaction, but it may come down to batching it and
> coding things such that I can catch and discard the offending row and retry
> the transaction if it fails (which should be fairly rare I would hope). I

Don't really know if it'd be better, or worse, or what, but another
thought to throw out there is to perhaps use savepoints instead of full
transactions?  Not sure if that's more expensive or cheaper than doing
full commits but it might be something to consider.

> When I finally get all of this sorted out and working reasonably optimally,
> I'll be sure to come back and report what techniques/settings did and didn't
> work for this workload.

That'd be great, many thanks,

    Stephen

On Mon, 12 Sep 2005 16:04:06 -0500
Brandon Black <blblack@gmail.com> wrote:

I've seen serveral tests PostgreSQL on JFS file system, it runs faster than using ext3.
Our production server works using JFS and RAID10,
we have 250K+ transactions per day and everything is OK.
Try switching to separate RAID10 arrays and use one for xlog, and others for data.
If you are using indexes, try to put them on separate RAID10 array.

> Test hardware right now is a dual Opteron with 4G of ram, which we've barely
> gotten 1,000 clients running against. Current disk hardware in testing is
> whatever we could scrape together (4x 3-ware PCI hardware RAID controllers,
> with 8 SATA drives in a RAID10 array off of each - aggregated up in a 4-way
> stripe with linux md driver and then formatted as ext3 with an appropriate
> stride parameter and data=writeback). Production will hopefully be a 4-8-way
> Opteron, 16 or more G of RAM, and a fiberchannel hardware raid array or two
> (~ 1TB available RAID10 storage) with 15krpm disks and battery-backed write
> cache.

--
Evgeny Gridasov
Software Developer
I-Free, Russia

On Tue, 2005-09-13 at 11:30, Brandon Black wrote:
> I started looking closer at my options for partitioning (inheritance,
> union all), and at Bizgres today.  Bizgres partitioning appears to be
> basically the same kind of inheritance partitioning one can do in
> mainline PostgreSQL.  Am I correct in thinking that the main
> difference is that they've coded support for
> "enable_constraint_exclusion=true" so that the query planner can be
> more effective at taking advantage of the partitioning when you've
> specified CHECK constraints on the child tables?  I may go for 8.1
> instead in that case, as the main win I'm looking for is that with
> inheritance I'll be doing inserts into smaller tables instead of ones
> that grow to unmanageable sizes (and that I can drop old child tables
> instead of delete/vacuum).

Perhaps I missed something in this thread, but don't forget
you still need vacuum to reclaim XIDs.

    --Ian

On Tue, Sep 13, 2005 at 12:16:55PM -0400, Ian Westmacott wrote:
> On Tue, 2005-09-13 at 11:30, Brandon Black wrote:
> > I started looking closer at my options for partitioning (inheritance,
> > union all), and at Bizgres today.  Bizgres partitioning appears to be
> > basically the same kind of inheritance partitioning one can do in
> > mainline PostgreSQL.  Am I correct in thinking that the main
> > difference is that they've coded support for
> > "enable_constraint_exclusion=true" so that the query planner can be
> > more effective at taking advantage of the partitioning when you've
> > specified CHECK constraints on the child tables?  I may go for 8.1
> > instead in that case, as the main win I'm looking for is that with
> > inheritance I'll be doing inserts into smaller tables instead of ones
> > that grow to unmanageable sizes (and that I can drop old child tables
> > instead of delete/vacuum).
>
> Perhaps I missed something in this thread, but don't forget
> you still need vacuum to reclaim XIDs.

Yes, but if you are going to drop the partition before 1 billion
transactions, you can skip vacuuming it completely.

--
Alvaro Herrera -- Valdivia, Chile         Architect, www.EnterpriseDB.com
"Es filósofo el que disfruta con los enigmas" (G. Coli)