In order to have the option to use O_DIRECT/FILE_FLAG_NO_BUFFERING in a
later commit, we need the addresses of user space buffers to be well
aligned. The exact requirements vary by OS and file system (typically
sectors and/or memory pages). The address alignment size is set to
4096, which is enough for currently known systems: it matches modern
sectors and common memory page size. There is no standard governing
O_DIRECT's requirements so we might eventually have to reconsider this
with more information from the field or future systems.
Aligning I/O buffers on memory pages is also known to improve regular
buffered I/O performance.
Three classes of I/O buffers for regular data pages are adjusted:
(1) Heap buffers are now allocated with the new palloc_aligned() or
MemoryContextAllocAligned() functions introduced by commit 439f6175.
(2) Stack buffers now use a new struct PGIOAlignedBlock to respect
PG_IO_ALIGN_SIZE, if possible with this compiler. (3) The buffer
pool is also aligned in shared memory.
WAL buffers were already aligned on XLOG_BLCKSZ. It's possible for
XLOG_BLCKSZ to be configured smaller than PG_IO_ALIGNED_SIZE and thus
for O_DIRECT WAL writes to fail to be well aligned, but that's a
pre-existing condition and will be addressed by a later commit.
BufFiles are not yet addressed (there's no current plan to use O_DIRECT
for those, but they could potentially get some incidental speedup even
in plain buffered I/O operations through better alignment).
If we can't align stack objects suitably using the compiler extensions
we know about, we disable the use of O_DIRECT by setting PG_O_DIRECT to
0. This avoids the need to consider systems that have O_DIRECT but
can't align stack objects the way we want; such systems could in theory
be supported with more work but we don't currently know of any such
machines, so it's easier to pretend there is no O_DIRECT support
instead. That's an existing and tested class of system.
Add assertions that all buffers passed into smgrread(), smgrwrite() and
smgrextend() are correctly aligned, unless PG_O_DIRECT is 0 (= stack
alignment tricks may be unavailable) or the block size has been set too
small to allow arrays of buffers to be all aligned.
Author: Thomas Munro <thomas.munro@gmail.com>
Author: Andres Freund <andres@anarazel.de>
Reviewed-by: Justin Pryzby <pryzby@telsasoft.com>
Discussion: https://postgr.es/m/CA+hUKGK1X532hYqJ_MzFWt0n1zt8trz980D79WbjwnT-yYLZpg@mail.gmail.com
Instead of referring to target backends by pid, use pgprocno. This
means that we don't have to scan the ProcArray and we can drop some
special case code for dealing with the startup process.
Discussion: https://postgr.es/m/CA%2BhUKGLYRyDaneEwz5Uya_OgFLMx5BgJfkQSD%3Dq9HmwsfRRb-w%40mail.gmail.com
Reviewed-by: Soumyadeep Chakraborty <soumyadeep2007@gmail.com>
Reviewed-by: Ashwin Agrawal <ashwinstar@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
1. Backends waiting for buffer I/O are now interruptible.
2. If something goes wrong in a backend that is currently performing
I/O, waiting backends no longer wake up until that backend reaches
AbortBufferIO() and broadcasts on the CV. Previously, any waiters would
wake up (because the I/O lock was automatically released) and then
busy-loop until AbortBufferIO() cleared BM_IO_IN_PROGRESS.
3. LWLockMinimallyPadded is removed, as it would now be unused.
Author: Robert Haas <robertmhaas@gmail.com>
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Reviewed-by: Julien Rouhaud <rjuju123@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us> (earlier version, 2016)
Discussion: https://postgr.es/m/CA%2BhUKGJ8nBFrjLuCTuqKN0pd2PQOwj9b_jnsiGFFMDvUxahj_A%40mail.gmail.com
Discussion: https://postgr.es/m/CA+Tgmoaj2aPti0yho7FeEf2qt-JgQPRWb0gci_o1Hfr=C56Xng@mail.gmail.com
Choose names that fit into the conventions for wait event names
(particularly, that multi-word names are in the style MultiWordName)
and hopefully convey more information to non-hacker users than the
previous names did.
Also rename SerializablePredicateLockListLock to
SerializablePredicateListLock; the old name was long enough to cause
table formatting problems, plus the double occurrence of "Lock" seems
confusing/error-prone.
Also change a couple of particularly opaque LWLock field names.
Discussion: https://postgr.es/m/28683.1589405363@sss.pgh.pa.us
There is little point in using the LWLockRegisterTranche mechanism for
built-in tranche names. It wastes cycles, it creates opportunities for
bugs (since failing to register a tranche name is a very hard-to-detect
problem), and the lack of any centralized list of names encourages
sloppy nonconformity in name choices. Moreover, since we have a
centralized list of the tranches anyway in enum BuiltinTrancheIds, we're
certainly not buying any flexibility in return for these disadvantages.
Hence, nuke all the backend-internal LWLockRegisterTranche calls,
and instead provide a const array of the builtin tranche names.
(I have in mind to change a bunch of these names shortly, but this
patch is just about getting them into one place.)
Discussion: https://postgr.es/m/9056.1589419765@sss.pgh.pa.us
array_base and array_stride were added so that we could identify the
offset of an LWLock within a tranche, but this facility is only very
marginally used apart from the main tranche. So, give every lock in
the main tranche its own tranche ID and get rid of array_base,
array_stride, and all that's attached. For debugging facilities
(Trace_lwlocks and LWLOCK_STATS) print the pointer address of the
LWLock using %p instead of the offset. This is arguably more useful,
and certainly a lot cheaper. Drop the offset-within-tranche from
the information reported to dtrace and from one can't-happen message
inside lwlock.c.
The main user-visible impact of this change is that pg_stat_activity
will now report all waits for LWLocks as "LWLock" rather than
reporting some as "LWLockTranche" and others as "LWLockNamed".
The main motivation for this change is that the need to specify an
array_base and an array_stride is awkward for parallel query. There
is only a very limited supply of tranche IDs so we can't just keep
allocating new ones, and if we try to use the same tranche IDs every
time then we run into trouble when multiple parallel contexts are
use simultaneously. So if we didn't get rid of this mechanism we'd
have to make it even more complicated. By simplifying it in this
way, we instead reduce the size of the generated code for lwlock.c
by about 5%.
Discussion: http://postgr.es/m/CA+TgmoYsFn6NUW1x0AZtupJGUAs1UDY4dJtCN47_Q6D0sP80PA@mail.gmail.com
That commit increased all shared memory allocations to the next higher
multiple of PG_CACHE_LINE_SIZE, but it didn't ensure that allocation
started on a cache line boundary. It also failed to remove a couple
other pieces of now-useless code.
BUFFERALIGN() is perhaps obsolete at this point, and likely should be
removed at some point, too, but that seems like it can be left to a
future cleanup.
Mistakes all pointed out by Andres Freund. The patch is mine, with
a few extra assertions which I adopted from his version of this fix.
Pinning/Unpinning a buffer is a very frequent operation; especially in
read-mostly cache resident workloads. Benchmarking shows that in various
scenarios the spinlock protecting a buffer header's state becomes a
significant bottleneck. The problem can be reproduced with pgbench -S on
larger machines, but can be considerably worse for queries which touch
the same buffers over and over at a high frequency (e.g. nested loops
over a small inner table).
To allow atomic operations to be used, cram BufferDesc's flags,
usage_count, buf_hdr_lock, refcount into a single 32bit atomic variable;
that allows to manipulate them together using 32bit compare-and-swap
operations. This requires reducing MAX_BACKENDS to 2^18-1 (which could
be lifted by using a 64bit field, but it's not a realistic configuration
atm).
As not all operations can easily implemented in a lockfree manner,
implement the previous buf_hdr_lock via a flag bit in the atomic
variable. That way we can continue to lock the header in places where
it's needed, but can get away without acquiring it in the more frequent
hot-paths. There's some additional operations which can be done without
the lock, but aren't in this patch; but the most important places are
covered.
As bufmgr.c now essentially re-implements spinlocks, abstract the delay
logic from s_lock.c into something more generic. It now has already two
users, and more are coming up; there's a follupw patch for lwlock.c at
least.
This patch is based on a proof-of-concept written by me, which Alexander
Korotkov made into a fully working patch; the committed version is again
revised by me. Benchmarking and testing has, amongst others, been
provided by Dilip Kumar, Alexander Korotkov, Robert Haas.
On a large x86 system improvements for readonly pgbench, with a high
client count, of a factor of 8 have been observed.
Author: Alexander Korotkov and Andres Freund
Discussion: 2400449.GjM57CE0Yg@dinodell
Up to now checkpoints were written in the order they're in the
BufferDescriptors. That's nearly random in a lot of cases, which
performs badly on rotating media, but even on SSDs it causes slowdowns.
To avoid that, sort checkpoints before writing them out. We currently
sort by tablespace, relfilenode, fork and block number.
One of the major reasons that previously wasn't done, was fear of
imbalance between tablespaces. To address that balance writes between
tablespaces.
The other prime concern was that the relatively large allocation to sort
the buffers in might fail, preventing checkpoints from happening. Thus
pre-allocate the required memory in shared memory, at server startup.
This particularly makes it more efficient to have checkpoint flushing
enabled, because that'll often result in a lot of writes that can be
coalesced into one flush.
Discussion: alpine.DEB.2.10.1506011320000.28433@sto
Author: Fabien Coelho and Andres Freund
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
Move the content lock directly into the BufferDesc, so that locking and
pinning a buffer touches only one cache line rather than two. Adjust
the definition of BufferDesc slightly so that this doesn't make the
BufferDesc any larger than one cache line (at least on platforms where
a spinlock is only 1 or 2 bytes).
We can't fit the I/O locks into the BufferDesc and stay within one
cache line, so move those to a completely separate tranche. This
leaves a relatively limited number of LWLocks in the main tranche, so
increase the padding of those remaining locks to a full cache line,
rather than allowing adjacent locks to share a cache line, hopefully
reducing false sharing.
Performance testing shows that these changes make little difference
on laptop-class machines, but help significantly on larger servers,
especially those with more than 2 sockets.
Andres Freund, originally based on an earlier patch by Simon Riggs.
Review and cosmetic adjustments (including heavy rewriting of the
comments) by me.
Use "a" and "an" correctly, mostly in comments. Two error messages were
also fixed (they were just elogs, so no translation work required). Two
function comments in pg_proc.h were also fixed. Etsuro Fujita reported one
of these, but I found a lot more with grep.
Also fix a few other typos spotted while grepping for the a/an typos.
For example, "consists out of ..." -> "consists of ...". Plus a "though"/
"through" mixup reported by Euler Taveira.
Many of these typos were in old code, which would be nice to backpatch to
make future backpatching easier. But much of the code was new, and I didn't
feel like crafting separate patches for each branch. So no backpatching.
Benchmarks has shown that aligning the buffer descriptor array to
cache lines is important for scalability; especially on bigger,
multi-socket, machines.
Currently the array sometimes already happens to be aligned by
happenstance, depending how large previous shared memory allocations
were. That can lead to wildly varying performance results after minor
configuration changes.
In addition to aligning the start of descriptor array, also force the
size of individual descriptors to be of a common cache line size (64
bytes). That happens to already be the case on 64bit platforms, but
this way we can change the struct BufferDesc more easily.
As the alignment primarily matters in highly concurrent workloads
which probably all are 64bit these days, and the space wastage of
element alignment would be a bit more noticeable on 32bit systems, we
don't force the stride to be cacheline sized on 32bit platforms for
now. If somebody does actual performance testing, we can reevaluate
that decision by changing the definition of BUFFERDESC_PADDED_SIZE.
Discussion: 20140202151319.GD32123@awork2.anarazel.de
Per discussion with Bruce Momjan, Tom Lane, Robert Haas, and Peter
Geoghegan.
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
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.
to 'Size' (that is, size_t), and install overflow detection checks in it.
This allows us to remove the former arbitrary restrictions on NBuffers
etc. It won't make any difference in a 32-bit machine, but in a 64-bit
machine you could theoretically have terabytes of shared buffers.
(How efficiently we could manage 'em remains to be seen.) Similarly,
num_temp_buffers, work_mem, and maintenance_work_mem can be set above
2Gb on a 64-bit machine. Original patch from Koichi Suzuki, additional
work by moi.
computation. On modern machines this is as fast if not faster, and we
don't have to clog the CPU's L2 cache with a tens-of-KB pointer array.
If we ever decide to adopt a more dynamic allocation method for shared
buffers, we'll probably have to revert this patch, but in the meantime
we might as well save a few bytes and nanoseconds. Per Qingqing Zhou.
exit, instead of trying to take shortcuts. Introduce some additional
shutdown callback routines to eliminate kluges like having ProcKill
be responsible for shutting down the buffer manager. Ensure that the
order of operations during shutdown is predictable and what you would
expect given the module layering.
communication structure, and make it its own module with its own lock.
This should reduce contention at least a little, and it definitely makes
the code seem cleaner. Per my recent proposal.
the freelist, plus per-buffer spinlocks that protect access to individual
shared buffer headers. This requires abandoning a global freelist (since
the freelist is a global contention point), which shoots down ARC and 2Q
as well as plain LRU management. Adopt a clock sweep algorithm instead.
Preliminary results show substantial improvement in multi-backend situations.
Also performed an initial run through of upgrading our Copyright date to
extend to 2005 ... first run here was very simple ... change everything
where: grep 1996-2004 && the word 'Copyright' ... scanned through the
generated list with 'less' first, and after, to make sure that I only
picked up the right entries ...
Thomas Munro
Bruce Momjian
Robert Haas
Tom Lane
Amit Kapila
Michael Paquier
Andres Freund
Heikki Linnakangas
Magnus Hagander
PostgreSQL Daemon