While beneficial, both for throughput and average/worst case latency, in
a significant number of workloads, there are other workloads in which
backend_flush_after can cause significant performance regressions in
comparison to < 9.6 releases. The regression is most likely when the hot
data set is bigger than shared buffers, but significantly smaller than
the operating system's page cache.
I personally think that the benefit of enabling backend flush control is
considerably bigger than the potential downsides, but a fair argument
can be made that not regressing is more important than improving
performance/latency. As the latter is the consensus, change the default
to 0.
The other settings introduced in 428b1d6b2 do not have the same
potential for regressions, so leave them enabled.
Benchmarks leading up to changing the default have been performed by
Mithun Cy, Ashutosh Sharma and Robert Haas.
Discussion: CAD__OuhPmc6XH=wYRm_+Q657yQE88DakN4=Ybh2oveFasHkoeA@mail.gmail.com
Even with old_snapshot_threshold = -1 (which disables the "snapshot
too old" feature), performance regressions were seen at moderate to
high concurrency. For example, a one-socket, four-core system
running 200 connections at saturation could see up to a 2.3%
regression, with larger regressions possible on NUMA machines.
By inlining the early (smaller, faster) tests in the
TestForOldSnapshot() function, the i7 case dropped to a 0.2%
regression, which could easily just be noise, and is clearly an
improvement. Further testing will show whether more is needed.
The reverted changes were intended to force a choice of whether any
newly-added BufferGetPage() calls needed to be accompanied by a
test of the snapshot age, to support the "snapshot too old"
feature. Such an accompanying test is needed in about 7% of the
cases, where the page is being used as part of a scan rather than
positioning for other purposes (such as DML or vacuuming). The
additional effort required for back-patching, and the doubt whether
the intended benefit would really be there, have indicated it is
best just to rely on developers to do the right thing based on
comments and existing usage, as we do with many other conventions.
This change should have little or no effect on generated executable
code.
Motivated by the back-patching pain of Tom Lane and Robert Haas
pg_xlogdump includes bufmgr.h. With a compiler that emits code for
static inline functions even when they're unreferenced, that leads
to unresolved external references in the new static-inline version
of BufferGetPage(). So hide it with #ifndef FRONTEND, as we've done
for similar issues elsewhere. Per buildfarm member pademelon.
I was initially concerned that the some of the hundreds of
references to BufferGetPage() where the literal
BGP_NO_SNAPSHOT_TEST were passed might not optimize as well as a
macro, leading to some hard-to-find performance regressions in
corner cases. Inspection of disassembled code has shown identical
code at all inspected locations, and the size difference doesn't
amount to even one byte per such call. So make it readable.
Per gripes from Álvaro Herrera and Tom Lane
This feature is controlled by a new old_snapshot_threshold GUC. A
value of -1 disables the feature, and that is the default. The
value of 0 is just intended for testing. Above that it is the
number of minutes a snapshot can reach before pruning and vacuum
are allowed to remove dead tuples which the snapshot would
otherwise protect. The xmin associated with a transaction ID does
still protect dead tuples. A connection which is using an "old"
snapshot does not get an error unless it accesses a page modified
recently enough that it might not be able to produce accurate
results.
This is similar to the Oracle feature, and we use the same SQLSTATE
and error message for compatibility.
This patch is a no-op patch which is intended to reduce the chances
of failures of omission once the functional part of the "snapshot
too old" patch goes in. It adds parameters for snapshot, relation,
and an enum to specify whether the snapshot age check needs to be
done for the page at this point. This initial patch passes NULL
for the first two new parameters and BGP_NO_SNAPSHOT_TEST for the
third. The follow-on patch will change the places where the test
needs to be made.
Currently writes to the main data files of postgres all go through the
OS page cache. This means that some operating systems can end up
collecting a large number of dirty buffers in their respective page
caches. When these dirty buffers are flushed to storage rapidly, be it
because of fsync(), timeouts, or dirty ratios, latency for other reads
and writes can increase massively. This is the primary reason for
regular massive stalls observed in real world scenarios and artificial
benchmarks; on rotating disks stalls on the order of hundreds of seconds
have been observed.
On linux it is possible to control this by reducing the global dirty
limits significantly, reducing the above problem. But global
configuration is rather problematic because it'll affect other
applications; also PostgreSQL itself doesn't always generally want this
behavior, e.g. for temporary files it's undesirable.
Several operating systems allow some control over the kernel page
cache. Linux has sync_file_range(2), several posix systems have msync(2)
and posix_fadvise(2). sync_file_range(2) is preferable because it
requires no special setup, whereas msync() requires the to-be-flushed
range to be mmap'ed. For the purpose of flushing dirty data
posix_fadvise(2) is the worst alternative, as flushing dirty data is
just a side-effect of POSIX_FADV_DONTNEED, which also removes the pages
from the page cache. Thus the feature is enabled by default only on
linux, but can be enabled on all systems that have any of the above
APIs.
While desirable and likely possible this patch does not contain an
implementation for windows.
With the infrastructure added, writes made via checkpointer, bgwriter
and normal user backends can be flushed after a configurable number of
writes. Each of these sources of writes controlled by a separate GUC,
checkpointer_flush_after, bgwriter_flush_after and backend_flush_after
respectively; they're separate because the number of flushes that are
good are separate, and because the performance considerations of
controlled flushing for each of these are different.
A later patch will add checkpoint sorting - after that flushes from the
ckeckpoint will almost always be desirable. Bgwriter flushes are most of
the time going to be random, which are slow on lots of storage hardware.
Flushing in backends works well if the storage and bgwriter can keep up,
but if not it can have negative consequences. This patch is likely to
have negative performance consequences without checkpoint sorting, but
unfortunately so has sorting without flush control.
Discussion: alpine.DEB.2.10.1506011320000.28433@sto
Author: Fabien Coelho and Andres Freund
At the end of crash recovery, unlogged relations are reset to the empty
state, using their init fork as the template. The init fork is copied to
the main fork without going through shared buffers. Unfortunately WAL
replay so far has not necessarily flushed writes from shared buffers to
disk at that point. In normal crash recovery, and before the
introduction of 'fast promotions' in fd4ced523 / 9.3, the
END_OF_RECOVERY checkpoint flushes the buffers out in time. But with
fast promotions that's not the case anymore.
To fix, force WAL writes targeting the init fork to be flushed
immediately (using the new FlushOneBuffer() function). In 9.5+ that
flush can centrally be triggered from the code dealing with restoring
full page writes (XLogReadBufferForRedoExtended), in earlier releases
that responsibility is in the hands of XLOG_HEAP_NEWPAGE's replay
function.
Backpatch to 9.1, even if this currently is only known to trigger in
9.3+. Flushing earlier is more robust, and it is advantageous to keep
the branches similar.
Typical symptoms of this bug are errors like
'ERROR: index "..." contains unexpected zero page at block 0'
shortly after promoting a node.
Reported-By: Thom Brown
Author: Andres Freund and Michael Paquier
Discussion: 20150326175024.GJ451@alap3.anarazel.de
Backpatch: 9.1-
Per discussion, nowadays it is possible to have tablespaces that have
wildly different I/O characteristics from others. Setting different
effective_io_concurrency parameters for those has been measured to
improve performance.
Author: Julien Rouhaud
Reviewed by: Andres Freund
There was a window in RestoreBackupBlock where a page would be zeroed out,
but not yet locked. If a backend pinned and locked the page in that window,
it saw the zeroed page instead of the old page or new page contents, which
could lead to missing rows in a result set, or errors.
To fix, replace RBM_ZERO with RBM_ZERO_AND_LOCK, which atomically pins,
zeroes, and locks the page, if it's not in the buffer cache already.
In stable branches, the old RBM_ZERO constant is renamed to RBM_DO_NOT_USE,
to avoid breaking any 3rd party extensions that might use RBM_ZERO. More
importantly, this avoids renumbering the other enum values, which would
cause even bigger confusion in extensions that use ReadBufferExtended, but
haven't been recompiled.
Backpatch to all supported versions; this has been racy since hot standby
was introduced.
Since the dawn of time (aka Postgres95) multiple pins of the same
buffer by one backend have been optimized not to modify the shared
refcount more than once. This optimization has always used a NBuffer
sized array in each backend keeping track of a backend's pins.
That array (PrivateRefCount) was one of the biggest per-backend memory
allocations, depending on the shared_buffers setting. Besides the
waste of memory it also has proven to be a performance bottleneck when
assertions are enabled as we make sure that there's no remaining pins
left at the end of transactions. Also, on servers with lots of memory
and a correspondingly high shared_buffers setting the amount of random
memory accesses can also lead to poor cpu cache efficiency.
Because of these reasons a backend's buffers pins are now kept track
of in a small statically sized array that overflows into a hash table
when necessary. Benchmarks have shown neutral to positive performance
results with considerably lower memory usage.
Patch by me, review by Robert Haas.
Discussion: 20140321182231.GA17111@alap3.anarazel.de
In ordinary operation, VACUUM must be careful to take a cleanup lock on
each leaf page of a btree index; this ensures that no indexscans could
still be "in flight" to heap tuples due to be deleted. (Because of
possible index-tuple motion due to concurrent page splits, it's not enough
to lock only the pages we're deleting index tuples from.) In Hot Standby,
the WAL replay process must likewise lock every leaf page. There were
several bugs in the code for that:
* The replay scan might come across unused, all-zero pages in the index.
While btree_xlog_vacuum itself did the right thing (ie, nothing) with
such pages, xlogutils.c supposed that such pages must be corrupt and
would throw an error. This accounts for various reports of replication
failures with "PANIC: WAL contains references to invalid pages". To
fix, add a ReadBufferMode value that instructs XLogReadBufferExtended
not to complain when we're doing this.
* btree_xlog_vacuum performed the extra locking if standbyState ==
STANDBY_SNAPSHOT_READY, but that's not the correct test: we won't open up
for hot standby queries until the database has reached consistency, and
we don't want to do the extra locking till then either, for fear of reading
corrupted pages (which bufmgr.c would complain about). Fix by exporting a
new function from xlog.c that will report whether we're actually in hot
standby replay mode.
* To ensure full coverage of the index in the replay scan, btvacuumscan
would emit a dummy WAL record for the last page of the index, if no
vacuuming work had been done on that page. However, if the last page
of the index is all-zero, that would result in corruption of said page,
since the functions called on it weren't prepared to handle that case.
There's no need to lock any such pages, so change the logic to target
the last normal leaf page instead.
The first two of these bugs were diagnosed by Andres Freund, the other one
by me. Fixes based on ideas from Heikki Linnakangas and myself.
This has been wrong since Hot Standby was introduced, so back-patch to 9.0.
MarkBufferDirtyHint() writes WAL, and should know if it's got a
standard buffer or not. Currently, the only callers where buffer_std
is false are related to the FSM.
In passing, rename XLOG_HINT to XLOG_FPI, which is more descriptive.
Back-patch to 9.3.
Checksums are set immediately prior to flush out of shared buffers
and checked when pages are read in again. Hint bit setting will
require full page write when block is dirtied, which causes various
infrastructure changes. Extensive comments, docs and README.
WARNING message thrown if checksum fails on non-all zeroes page;
ERROR thrown but can be disabled with ignore_checksum_failure = on.
Feature enabled by an initdb option, since transition from option off
to option on is long and complex and has not yet been implemented.
Default is not to use checksums.
Checksum used is WAL CRC-32 truncated to 16-bits.
Simon Riggs, Jeff Davis, Greg Smith
Wide input and assistance from many community members. Thank you.
When relations are dropped, at end of transaction we need to remove the
files and clean the buffer pool of buffers containing pages of those
relations. Previously we would scan the buffer pool once per relation
to clean up buffers. When there are many relations to drop, the
repeated scans make this process slow; so we now instead pass a list of
relations to drop and scan the pool once, checking each buffer against
the passed list. When the number of relations is larger than a
threshold (which as of this patch is being set to 20 relations) we sort
the array before starting, and bsearch the array; when it's smaller, we
simply scan the array linearly each time, because that's faster. The
exact optimal threshold value depends on many factors, but the
difference is not likely to be significant enough to justify making it
user-settable.
This has been measured to be a significant win (a 15x win when dropping
100,000 relations; an extreme case, but reportedly a real one).
Author: Tomas Vondra, some tweaks by me
Reviewed by: Robert Haas, Shigeru Hanada, Andres Freund, Álvaro Herrera
This provides a speedup of about 4X when NBuffers is large enough.
There is also a useful reduction in sinval traffic, since we
only do CacheInvalidateSmgr() once not once per fork.
Simon Riggs, reviewed and somewhat revised by Tom Lane
Currently, the only way to see the numbers this gathers is via
EXPLAIN (ANALYZE, BUFFERS), but the plan is to add visibility through
the stats collector and pg_stat_statements in subsequent patches.
Ants Aasma, reviewed by Greg Smith, with some further changes by me.
To make it wake up promptly when activity starts again, backends nudge it
by setting a latch in MarkBufferDirty(). The latch is kept set while
bgwriter is active, so there is very little overhead from that when the
system is busy. It is only armed before going into longer sleep.
Peter Geoghegan, with some changes by me.
We need not wait until the commit record is durably on disk, because
in the event of a crash the page we're updating with hint bits will
be gone anyway. Per off-list report from Heikki Linnakangas, this
can significantly degrade the performance of unlogged tables; I was
able to show a 2x speedup from this patch on a pgbench run with scale
factor 15. In practice, this will mostly help small, heavily updated
tables, because on larger tables you're unlikely to run into the same
row again before the commit record makes it out to disk.
The contents of an unlogged table are WAL-logged; thus, they are not
available on standby servers and are truncated whenever the database
system enters recovery. Indexes on unlogged tables are also unlogged.
Unlogged GiST indexes are not currently supported.
This allows us to reliably remove all leftover temporary relation
files on cluster startup without reference to system catalogs or WAL;
therefore, we no longer include temporary relations in XLOG_XACT_COMMIT
and XLOG_XACT_ABORT WAL records.
Since these changes require including a backend ID in each
SharedInvalSmgrMsg, the size of the SharedInvalidationMessage.id
field has been reduced from two bytes to one, and the maximum number
of connections has been reduced from INT_MAX / 4 to 2^23-1. It would
be possible to remove these restrictions by increasing the size of
SharedInvalidationMessage by 4 bytes, but right now that doesn't seem
like a good trade-off.
Review by Jaime Casanova and Tom Lane.
woken by alarm we send SIGUSR1 to all backends requesting that they
check to see if they are blocking Startup process. If so, they throw
ERROR/FATAL as for other conflict resolutions. Deadlock stop gap
removed. max_standby_delay = -1 option removed to prevent deadlock.
This patch also removes buffer-usage statistics from the track_counts
output, since this (or the global server statistics) is deemed to be a better
interface to this information.
Itagaki Takahiro, reviewed by Euler Taveira de Oliveira.
GUC variable effective_io_concurrency controls how many concurrent block
prefetch requests will be issued.
(The best way to handle this for plain index scans is still under debate,
so that part is not applied yet --- tgl)
Greg Stark
truncations in FSM code, call FreeSpaceMapTruncateRel from smgr_redo. To
make that cleaner from modularity point of view, move the WAL-logging one
level up to RelationTruncate, and move RelationTruncate and all the
related WAL-logging to new src/backend/catalog/storage.c file. Introduce
new RelationCreateStorage and RelationDropStorage functions that are used
instead of calling smgrcreate/smgrscheduleunlink directly. Move the
pending rel deletion stuff from smgrcreate/smgrscheduleunlink to the new
functions. This leaves smgr.c as a thin wrapper around md.c; all the
transactional stuff is now in storage.c.
This will make it easier to add new forks with similar truncation logic,
like the visibility map.
(but not locked, as that would risk deadlocks). Also, make it work in a small
ring of buffers to avoid having bulk inserts trash the whole buffer arena.
Robert Haas, after an idea of Simon Riggs'.
functions into one ReadBufferExtended function, that takes the strategy
and mode as argument. There's three modes, RBM_NORMAL which is the default
used by plain ReadBuffer(), RBM_ZERO, which replaces ZeroOrReadBuffer, and
a new mode RBM_ZERO_ON_ERROR, which allows callers to read corrupt pages
without throwing an error. The FSM needs the new mode to recover from
corrupt pages, which could happend if we crash after extending an FSM file,
and the new page is "torn".
Add fork number to some error messages in bufmgr.c, that still lacked it.
of multiple forks, and each fork can be created and grown separately.
The bulk of this patch is about changing the smgr API to include an extra
ForkNumber argument in every smgr function. Also, smgrscheduleunlink and
smgrdounlink no longer implicitly call smgrclose, because other forks might
still exist after unlinking one. The callers of those functions have been
modified to call smgrclose instead.
This patch in itself doesn't have any user-visible effect, but provides the
infrastructure needed for upcoming patches. The additional forks envisioned
are a rewritten FSM implementation that doesn't rely on a fixed-size shared
memory block, and a visibility map to allow skipping portions of a table in
VACUUM that have no dead tuples.
corresponding struct definitions. This allows other headers to avoid including
certain highly-loaded headers such as rel.h and relscan.h, instead using just
relcache.h, heapam.h or genam.h, which are more lightweight and thus cause less
unnecessary dependencies.
forks. XLogOpenRelation() and the associated light-weight relation cache in
xlogutils.c is gone, and XLogReadBuffer() now takes a RelFileNode as argument,
instead of Relation.
For functions that still need a Relation struct during WAL replay, there's a
new function called CreateFakeRelcacheEntry() that returns a fake entry like
XLogOpenRelation() used to.
more logical that way, and also it reduces the amount of unnecessary includes
in bufpage.h, which is widely used.
Zdenek Kotala.
My previous patch to bufpage.h should also have credited him as author, but I
forgot (sorry about that).
buffers that cannot possibly need to be cleaned, and estimates how many
buffers it should try to clean based on moving averages of recent allocation
requests and density of reusable buffers. The patch also adds a couple
more columns to pg_stat_bgwriter to help measure the effectiveness of the
bgwriter.
Greg Smith, building on his own work and ideas from several other people,
in particular a much older patch from Itagaki Takahiro.
columns, and the new version can be stored on the same heap page, we no longer
generate extra index entries for the new version. Instead, index searches
follow the HOT-chain links to ensure they find the correct tuple version.
In addition, this patch introduces the ability to "prune" dead tuples on a
per-page basis, without having to do a complete VACUUM pass to recover space.
VACUUM is still needed to clean up dead index entries, however.
Pavan Deolasee, with help from a bunch of other people.
