Skip site navigation (1) Skip section navigation (2)

Re: Advice/guideline on increasing shared_buffers and kernel parameters

From: "Kevin Grittner" <Kevin(dot)Grittner(at)wicourts(dot)gov>
To: <pgsql-admin(at)postgresql(dot)org>,<gnanam(at)zoniac(dot)com>
Subject: Re: Advice/guideline on increasing shared_buffers and kernel parameters
Date: 2012-05-09 14:57:08
Message-ID: (view raw, whole thread or download thread mbox)
Lists: pgsql-admin
"Gnanakumar" <gnanam(at)zoniac(dot)com> wrote:
>> We get very good performance dealing with thousands of concurrent
>> users with a pool of 35 connections to the database.
>> If you want to handle more users than you can currently support,
>> you probably need to use fewer database connections.
> First, please excuse me that I'm not able to understand this
> particular point clearly. How can be reducing/using fewer
> connections in connection pooler can support larger concurrent
> incoming connection requests? If this is so critical to revisit
> (reducing), then I may have to convince/justify my peers also,
> before making this change in the Production server. Can you throw
> some light on this subject?
> Thanks for bringing this idea to notice.
There have been numerous discussions of this on the lists, so you
can probably find a more in-depth discussion of the topic if you
search the archives, and this may motivate me to put together a Wiki
page on the topic, but here's the general concept.
A database server only has so many resources, and if you don't have
enough active connections active to use all of them, your throughput
will generally improve by using more connections.  Once all of the
resources are in use, you won't push any more through by having more
connections competing for the resources.  In fact, throughput starts
to fall off due to the overhead from that contention.  If you look
at any graph of PostgreSQL performance with number of connections on
the x axis and tps on the y access (with nothing else changing), you
will performance climb as connections rise until you hit saturation,
and then you have a "knee" after which performance falls off.  A lot
of work has been done for version 9.3 to push that knee to the right
and make the fall-off more gradual, but the issue is intrinsic --
without a built-in connection pool or at least an admission control
policy, the knee will always be there.
Now, this decision not to include a connection pooler inside the
PostgreSQL server itself is not capricious and arbitrary.  In many
cases you will get better performance if the connection pooler is
running on a separate machine.  In even more cases (at least in my
experience) you can get improved functionality by incorporating a
connection pool into client-side software.  Many frameworks,
including the ones we use at Wisconsin Courts, do the pooling in a
Java process running on the same server as the database server (to
minimize latency effects from the database protocol) and make
high-level requests to the Java process to run a certain function
with a given set of parameters as a single database transaction. 
This ensures that network latency or connection failures can't cause
a transaction to hang while waiting for something from the network,
and provides a simple way to retry any database transaction which
rolls back with a serialization failure (SQLSTATE 40001 or 40P01).
Since a pooler built in to the database engine would be inferior
(for the above reasons), the community has decided not to go that
I know I won't be able to remember *all* of the reasons that
performance *falls off* after you reach the "knee" rather than just
staying level, but I'll list the ones which come to mind at the
moment.  If anyone wants to add to the list, feel free to reply, or
look for a Wiki page to appear this week and add them there.
 - Context switches.  The processor is interrupted from working on
one query and has to switch to another, which involves saving state
and restoring state.  While the core is busy swapping states it is
not doing any useful work on any query.
 - Cache line contention.  One query is likely to be working on a
particular area of RAM, and the query taking its place is likely to
be working on a different area; causing data cached on the CPU chip
to be discarded, only to need to be reloaded to continue the other
query.  Besides that the various processes will be grabbing control
of cache lines from each other, causing stalls.  (Humorous note, in
one oprofile run of a heavily contended load, 10% of CPU time was
attributed to a 1-byte noop; analysis showed that it was because it
needed to wait on a cache line for the following machine code
 - Lock contention.  This happens at various levels: spinlocks, LW
locks, and all the locks that show up in pg_locks.  As more
processes compete for the spinlocks (which protect LW locks
acquisition and release, which in turn protect the heavyweight and
predicate lock acquisition and release) they account for a high
percentage of CPU time used.
 - RAM usage.  The work_mem setting can have a big impact on
performance.  If it is too small, hash tables and sorts spill to
disk, bitmap heap scans become "lossy", requiring more work on each
page access, etc.  So you want it to be big.  But work_mem RAM can
be allocated for each node of a query on each connection, all at the
same time.  So a big work_mem with a large number of connections can
cause a lot of the OS cache to be periodically discarded, forcing
more accesses to disk; or it could even put the system into
swapping.  So the more connections you have, the more you need to
make a choice between slow plans and trashing cache/swapping.
 - Disk access.  If you *do* need to go to disk for random access, a
large number of connections can tend to force more tables and
indexes to be accessed at the same time, causing heavier seeking all
over the disk.
 - General scaling.  Some internal structures allocated based on
max_connections scale at O(N^2) or O(N*log(N)).  Some types of
overhead which are negligible at a lower number of connections can
become significant with a large number of connections.
A formula which has held up pretty well across a lot of benchmarks
for years is that for optimal throughput the number of active
connections should be somewhere near ((core_count * 2) +
effective_spindle_count).  Core count should not include HT threads,
even if hyperthreading is enabled.  Effective spindle count is zero
if the active data set is fully cached, and approaches the actual
number of spindles as the cache hit rate falls.  Benchmarks of WIP
for version 9.3 suggest that this formula will need adjustment on
that release.  I haven't looked at how well the formula works with
SDDs.  In any event, I would recommend using this as a starting
point for a connection pool size, and trying incremental adjustments
with your actual workload to find the actual "sweet spot" for your
hardware and workload.

In response to


pgsql-admin by date

Next:From: Kevin GrittnerDate: 2012-05-09 15:03:53
Subject: Re: FW: pg_dump: schema with OID 2200 does not exist
Previous:From: Elizandro GallegosDate: 2012-05-09 14:49:12
Subject: FW: pg_dump: schema with OID 2200 does not exist

Privacy Policy | About PostgreSQL
Copyright © 1996-2017 The PostgreSQL Global Development Group