DB cache size strategies

[Ed L.]

Here's some of my current notions on pgsql performance tuning strictly as it
relates to pgsql tuning parameters in the context of a dedicated linux or
hpux server. I'm particularly focusing on the shared_buffers setting. I
invite any corrective or confirming feedback. I realize there are many
other hugely important performance factors outside this scope.

One key aspect of pgsql performance tuning is to adjust the memory
consumption settings (shared_buffers , sort_mem, etc) large enough to hold
as much of the database in shared memory as possible while not causing page
swap-ins. I understand that both page swap-outs and swap space usage is
normal and OK, but lots of page swap-ins are bad). In other words, for
absolute fastest performance, we want a database buffer cache hit rate of
as close to 100% as possible.

Now, how do we accurately measure this cache hit rate at any one point?
Well, here's what I currently know as the best measure for a given cluster
of databases:

SELECT SUM(pg_stat_get _db_blocks_hit( d.oid))
/ SUM(pg_stat_get _db_blocks_fetc hed(d.oid)) AS hit_rate
FROM pg_database d;

How do we identify when we have sub-optimal pgsql configuration parameters
in this regard? My understanding: to whatever extent the hit rate is
below 100%, our shared buffers are not large enough to hold the data we're
accessing.

How do we fix the problem? We want to increase the shared_buffers setting
and/or maybe some combination sort_mem, wal_buffers, vacuum_mem, reset our
stats, and resume monitoring. I assume which of these we increase for
maximum benefit depends on the sources of the cache overflow (routine
queries, sorting, vacuuming), but that shared_buffers is the most impactful
in practice. Again, we want to do this without causing page swap-ins.
When you see swap-ins, you've gone too far. If we experience swap-ins and
less than 100% cache hit rate by any significant amount, then we need more
RAM or less RAM consumption.

The end result is that page swap-ins (sar -W) and cache hit rate (query
above) are two very key measures to use as warning bells.

Glad to hear your thoughts.

Ed
---------------------------(end of broadcast)---------------------------
TIP 2: you can get off all lists at once with the unregister command
(send "unregister YourEmailAddres sHere" to ma*******@postg resql.org)

[Ed L.]

On Friday January 30 2004 2:33, Ed L. wrote:

One key aspect of pgsql performance tuning is to adjust the memory
consumption settings (shared_buffers , sort_mem, etc) large enough to hold
as much of the database in shared memory as possible while not causing
page swap-ins. I understand that both page swap-outs and swap space
usage is normal and OK, but lots of page swap-ins are bad). In other
words, for absolute fastest performance, we want a database buffer cache
hit rate of as close to 100% as possible.

I'm also curious about the relationship of DB shared buffer cache to the
linux/hpux kernel caches. In particular, if the block being retrieved in
pgsql was in the kernel's cache but not in the DB cache, thereby forcing a
read() system call, what kind of quantitative difference in performance
would one expect when comparing with block retrievals coming from the
cache? I would think they'd differ only by something on the order of
microseconds. Is the linux kernel disk cache normally a duplicate of much
of what is in the DB cache? For linux, does the kernel cache use only
"available" memory until a program needs it, while the pgsql DB cache
memory is guaranteed at startup?

TIA.

Ed

---------------------------(end of broadcast)---------------------------
TIP 8: explain analyze is your friend

[Martijn van Oosterhout]

On Fri, Jan 30, 2004 at 03:19:56PM -0700, Ed L. wrote:

On Friday January 30 2004 2:33, Ed L. wrote:

One key aspect of pgsql performance tuning is to adjust the memory
consumption settings (shared_buffers , sort_mem, etc) large enough to hold
as much of the database in shared memory as possible while not causing
page swap-ins. I understand that both page swap-outs and swap space
usage is normal and OK, but lots of page swap-ins are bad). In other
words, for absolute fastest performance, we want a database buffer cache
hit rate of as close to 100% as possible.
I'm also curious about the relationship of DB shared buffer cache to the
linux/hpux kernel caches. In particular, if the block being retrieved in
pgsql was in the kernel's cache but not in the DB cache, thereby forcing a
read() system call, what kind of quantitative difference in performance
would one expect when comparing with block retrievals coming from the
cache? I would think they'd differ only by something on the order of
microseconds. Is the linux kernel disk cache normally a duplicate of much
of what is in the DB cache? For linux, does the kernel cache use only
"available" memory until a program needs it, while the pgsql DB cache
memory is guaranteed at startup?

Whenever the database needs a block not in memory it get loaded and ends up
in both the OS cache and the DB cache. The difference between getting a
block out of DB cache and OS cache is very, very small compared to loading
off disk.

Hence, the larger your DB cache, the less overall data you can cache due to
the duplication. OTOH, postgres can only directly use the DB cache and can
only access the OS cache indirectly.

What I try to aim for is the keep the DB cache somewhere near the "working
set", ie the amount of space used by all the system tables and commonly used
tables in the system. It's shared by all postgres servers in a cluster so
keep that in mind. Also the OS cache may be shared by other processes, so
your results will depend on what other things are happening.

Finally, at least on Linux, the shared memory postgres uses for cache can
also be swapped out making it very difficult to determine the correct value.

Hope this helps,

--
Martijn van Oosterhout <kl*****@svana. org> http://svana.org/kleptog/ (... have gone from d-i being barely usable even by its developers
anywhere, to being about 20% done. Sweet. And the last 80% usually takes
20% of the time, too, right?) -- Anthony Towns, debian-devel-announce

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.0.6 (GNU/Linux)
Comment: For info see http://www.gnupg.org

iD8DBQFAGv+nY5T wig3Ge+YRAlobAK DUzlZXSZt+tDn5c 3ewkSuj965lMACe KUT1
AV/p/m2Pvf8QUF3MTH/opXk=
=XaoN
-----END PGP SIGNATURE-----

[Jan Wieck]

But keep in mind that all these strategies need to be reevaluated with
7.5 and its ARC.
Jan

Martijn van Oosterhout wrote:

On Fri, Jan 30, 2004 at 03:19:56PM -0700, Ed L. wrote:

On Friday January 30 2004 2:33, Ed L. wrote:

>
> One key aspect of pgsql performance tuning is to adjust the memory
> consumption settings (shared_buffers , sort_mem, etc) large enough to hold
> as much of the database in shared memory as possible while not causing
> page swap-ins. I understand that both page swap-outs and swap space
> usage is normal and OK, but lots of page swap-ins are bad). In other
> words, for absolute fastest performance, we want a database buffer cache
> hit rate of as close to 100% as possible.

I'm also curious about the relationship of DB shared buffer cache to the
linux/hpux kernel caches. In particular, if the block being retrieved in
pgsql was in the kernel's cache but not in the DB cache, thereby forcing a
read() system call, what kind of quantitative difference in performance
would one expect when comparing with block retrievals coming from the
cache? I would think they'd differ only by something on the order of
microseconds. Is the linux kernel disk cache normally a duplicate of much
of what is in the DB cache? For linux, does the kernel cache use only
"available" memory until a program needs it, while the pgsql DB cache
memory is guaranteed at startup?

Whenever the database needs a block not in memory it get loaded and ends up
in both the OS cache and the DB cache. The difference between getting a
block out of DB cache and OS cache is very, very small compared to loading
off disk.

Hence, the larger your DB cache, the less overall data you can cache due to
the duplication. OTOH, postgres can only directly use the DB cache and can
only access the OS cache indirectly.

What I try to aim for is the keep the DB cache somewhere near the "working
set", ie the amount of space used by all the system tables and commonly used
tables in the system. It's shared by all postgres servers in a cluster so
keep that in mind. Also the OS cache may be shared by other processes, so
your results will depend on what other things are happening.

Finally, at least on Linux, the shared memory postgres uses for cache can
also be swapped out making it very difficult to determine the correct value.

Hope this helps,

--
#============== =============== =============== =============== ===========#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#============== =============== =============== ====== Ja******@Yahoo. com #
---------------------------(end of broadcast)---------------------------
TIP 2: you can get off all lists at once with the unregister command
(send "unregister YourEmailAddres sHere" to ma*******@postg resql.org)

[Ed L.]

On Friday January 30 2004 6:06, Martijn van Oosterhout wrote:

On Fri, Jan 30, 2004 at 03:19:56PM -0700, Ed L. wrote:

I'm also curious about the relationship of DB shared buffer cache to
the linux/hpux kernel caches. ...
Whenever the database needs a block not in memory it get loaded and ends
up in both the OS cache and the DB cache. The difference between getting
a block out of DB cache and OS cache is very, very small compared to
loading off disk.

So the OS cache and the DB cache may well have duplicate copies of the data?
With OS cache being more susceptible to preemption from other processes?
Finally, at least on Linux, the shared memory postgres uses for cache can
also be swapped out making it very difficult to determine the correct
value.

What would cause DB shared memory to be swapped out in linux? I thought it
was all pre-allocated at pgsql startup. Is this the "well its not really
pre-allocated but rather supplied on demand" story? I think I saw some
sort of kernel parameter controlling a similar (same?) policy...

TIA.

---------------------------(end of broadcast)---------------------------
TIP 4: Don't 'kill -9' the postmaster

[Martijn van Oosterhout]

On Tue, Feb 10, 2004 at 01:20:32PM -0700, Ed L. wrote:

On Friday January 30 2004 6:06, Martijn van Oosterhout wrote:

On Fri, Jan 30, 2004 at 03:19:56PM -0700, Ed L. wrote:

I'm also curious about the relationship of DB shared buffer cache to
the linux/hpux kernel caches. ...
Whenever the database needs a block not in memory it get loaded and ends
up in both the OS cache and the DB cache. The difference between getting
a block out of DB cache and OS cache is very, very small compared to
loading off disk.

So the OS cache and the DB cache may well have duplicate copies of the data?
With OS cache being more susceptible to preemption from other processes?

Will have. Any time postgresql does a read() the kernel is copying a buffer
from it's cache to postgresql's internal cache. The kernel cache is more
susceptable to reuse by other processes, but it's still double.

Finally, at least on Linux, the shared memory postgres uses for cache can
also be swapped out making it very difficult to determine the correct
value.

What would cause DB shared memory to be swapped out in linux? I thought it
was all pre-allocated at pgsql startup. Is this the "well its not really
pre-allocated but rather supplied on demand" story? I think I saw some
sort of kernel parameter controlling a similar (same?) policy...

Memory pressure. Same rules apply as for everything else I beleive. If you
allocate a huge amount to shared memory, fill it up and then don't use it
for a while while the rest of the system needs the memory, it'll get swapped
out.

I generally give Postgresql about 64-128MB of shared memory, which covers
all of the system tables and the most commonly used small tables. The rest
of the memory (this is a 1GB machine) I leave for the kernel to manage for
the very large tables.

I think giving postgresql half your memory is couterproductiv e.

--
Martijn van Oosterhout <kl*****@svana. org> http://svana.org/kleptog/ (... have gone from d-i being barely usable even by its developers
anywhere, to being about 20% done. Sweet. And the last 80% usually takes
20% of the time, too, right?) -- Anthony Towns, debian-devel-announce

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.0.6 (GNU/Linux)
Comment: For info see http://www.gnupg.org

iD8DBQFAKUI3Y5T wig3Ge+YRAjYrAJ 4zaoXQCtYI8tsXq XDl7W4tKIDtqgCg i7+u
Qa/QJnTIWaYt8IfdBh rKrnU=
=kntV
-----END PGP SIGNATURE-----

[Ed L.]

On Tuesday February 10 2004 1:42, Martijn van Oosterhout wrote:

I generally give Postgresql about 64-128MB of shared memory, which covers
all of the system tables and the most commonly used small tables. The
rest of the memory (this is a 1GB machine) I leave for the kernel to
manage for the very large tables.
Interesting. Why leave very large tables to the kernel instead of the db
cache? Assuming a dedicated DB server and a DB smaller than available RAM,
why not give the DB enough RAM to get the entire DB into the DB cache?
(Assuming you have the RAM).

Is there a point of diminishing returns for DB cache size *prior* to running
out of data to load into, and access from, that cache? Anyone know how to
guage that point? Maybe shared memory management overhead versus the
performance difference between DB cache reads and OS cache reads? We have
some pretty intense DB loads, but that feels like counting angels on the
head of a pin...
I think giving postgresql half your memory is couterproductiv e.

I get the sense it really depends on the nature of the apps running on the
box. If it's a dedicated server (not sharing with an app, a web server,
etc), then giving it enough to get the entire data in memory without
swapping is ideal (without swapping), no matter whether thats 10MB or 10GB
of DB cache. If its a cluster server running along side lots of other apps
and they all generally have the same priority, and assuming disk cache
reads are nearly as fast as DB cache reads, it seems the less DB cache you
have, the better; let all apps leverage the kernel cache equally. Would
you agree?

One troublesome case for which I wonder if there is an easy tweak
unbeknownst to me: a huge table that gets lots of random reads and writes.
It's too big for the cache, and the volume of random queries is such that
its a disk I/O drain. Of course, more memory or better data modeling could
help. Maybe that is the worst case scenario.

Thx.

---------------------------(end of broadcast)---------------------------
TIP 9: the planner will ignore your desire to choose an index scan if your
joining column's datatypes do not match

[scott.marlowe]

On Tue, 10 Feb 2004, Ed L. wrote:

On Tuesday February 10 2004 1:42, Martijn van Oosterhout wrote:

I generally give Postgresql about 64-128MB of shared memory, which covers
all of the system tables and the most commonly used small tables. The
rest of the memory (this is a 1GB machine) I leave for the kernel to
manage for the very large tables.

Interesting. Why leave very large tables to the kernel instead of the db
cache? Assuming a dedicated DB server and a DB smaller than available RAM,
why not give the DB enough RAM to get the entire DB into the DB cache?
(Assuming you have the RAM).

Because the kernel is more efficient (right now) at caching large data
sets.

With the ARC cache manager that will likely wend it's way into 7.5, it's
quite a likely possibility that postgresql will be able to efficiently
handle a larger cache, but it will still be a shared memory cache, and
those are still usually much slower than the kernel's cache.
---------------------------(end of broadcast)---------------------------
TIP 3: if posting/reading through Usenet, please send an appropriate
subscribe-nomail command to ma*******@postg resql.org so that your
message can get through to the mailing list cleanly

[Ed L.]

On Tuesday February 10 2004 3:48, scott.marlowe wrote:

On Tue, 10 Feb 2004, Ed L. wrote:

Interesting. Why leave very large tables to the kernel instead of the
db cache? Assuming a dedicated DB server and a DB smaller than
available RAM, why not give the DB enough RAM to get the entire DB into
the DB cache? (Assuming you have the RAM).

Because the kernel is more efficient (right now) at caching large data
sets.

With the ARC cache manager that will likely wend it's way into 7.5, it's
quite a likely possibility that postgresql will be able to efficiently
handle a larger cache, but it will still be a shared memory cache, and
those are still usually much slower than the kernel's cache.

Hmmm. Others have asserted/assumed they'd be roughly equivalent. It'd be
interesting to see some real data measuring the performance of the shared
mem cache vs. kernel cache. Anyone know of existing benchmarks?
---------------------------(end of broadcast)---------------------------
TIP 2: you can get off all lists at once with the unregister command
(send "unregister YourEmailAddres sHere" to ma*******@postg resql.org)

[scott.marlowe]

On Tue, 10 Feb 2004, Ed L. wrote:

On Tuesday February 10 2004 3:48, scott.marlowe wrote:

On Tue, 10 Feb 2004, Ed L. wrote:

Interesting. Why leave very large tables to the kernel instead of the
db cache? Assuming a dedicated DB server and a DB smaller than
available RAM, why not give the DB enough RAM to get the entire DB into
the DB cache? (Assuming you have the RAM).

Because the kernel is more efficient (right now) at caching large data
sets.

With the ARC cache manager that will likely wend it's way into 7.5, it's
quite a likely possibility that postgresql will be able to efficiently
handle a larger cache, but it will still be a shared memory cache, and
those are still usually much slower than the kernel's cache.

Hmmm. Others have asserted/assumed they'd be roughly equivalent. It'd be
interesting to see some real data measuring the performance of the shared
mem cache vs. kernel cache. Anyone know of existing benchmarks?

Testing from various folks has shown that there is a maximum size of
somewhere around 256 meg or so where the returns diminish pretty quickly,
unless your working data sets are larger than that. I.e. setting a
database up that needs 20 megs to throw data around in with 200 megs of
buffers is usually slower than handing it 32 megs of buffer and letting
the kernel do it.

But ARC may change all of that. I'll have to grab a CVS snapshot and see
how it behaves with large datasets and shared buffers...
---------------------------(end of broadcast)---------------------------
TIP 4: Don't 'kill -9' the postmaster