Jignesh Shah's scalability testing on Solaris has revealed further
tuning opportunities surrounding the start and end of a transaction.
Tuning that should be especially important since async commit is likely
to allow much higher transaction rates than were previously possible.
There is strong contention on the ProcArrayLock in Exclusive mode, with
the top path being CommitTransaction(). This becomes clear as the number
of connections increases, but it seems likely that the contention can be
caused in a range of other circumstances. My thoughts on the causes of
this contention are that the following 3 tasks contend with each other
in the following way:
CommitTransaction(): takes ProcArrayLock Exclusive
but only needs access to one ProcArray element
which waits for
takes XidGenLock Exclusive
ExtendCLOG(): takes ClogControlLock Exclusive, WALInsertLock Exclusive
two possible place where I/O is required
ExtendSubtrans(): takes SubtransControlLock()
one possible place where I/O is required
Avoids lock on ProcArrayLock: atomically updates one ProcArray element
or more simply:
CommitTransaction() -- i.e. once per transaction
GetSnapshotData() -- i.e. once per SQL statement
which waits for
GetNextTransactionId() -- i.e. once per transaction
This gives some goals for scalability improvements and some proposals.
(1) and (2) are proposals for 8.3 tuning, the others are directions for
Goal: Reduce total time that GetSnapshotData() waits for
1. Increase size of Clog-specific BLCKSZ
Clog currently uses BLCKSZ to define the size of clog buffers. This can
be changed to use CLOG_BLCKSZ, which would then be set to 32768.
This will naturally increase the amount of memory allocated to the clog,
so we need not alter CLOG_BUFFERS above 8 if we do this (as previously
suggested, with successful results). This will also reduce the number of
ExtendClog() calls, which will probably reduce the overall contention
2. Perform ExtendClog() as a background activity
Background process can look at the next transactionid once each cycle
without holding any lock. If the xid is almost at the point where a new
clog page would be allocated, then it will allocate one prior to the new
page being absolutely required. Doing this as a background task would
mean that we do not need to hold the XidGenLock in exclusive mode while
we do this, which means that GetSnapshotData() and CommitTransaction()
would also be less likely to block. Also, if any clog writes need to be
performed when the page is moved forwards this would also be performed
in the background.
3. Consider whether ProcArrayLock should use a new queued-shared lock
mode that puts a maximum wait time on ExclusiveLock requests. It would
be fairly hard to implement this well as a timer, but it might be
possible to place a limit on queue length. i.e. allow Share locks to be
granted immediately if a Shared holder already exists, but only if there
is a queue of no more than N exclusive mode requests queued. This might
prevent the worst cases of exclusive lock starvation.
4. Since shared locks are currently queued behind exclusive requests
when they cannot be immediately satisfied, it might be worth
reconsidering the way LWLockRelease works also. When we wake up the
queue we only wake the Shared requests that are adjacent to the head of
the queue. Instead we could wake *all* waiting Shared requestors.
e.g. with a lock queue like this:
Currently we would wake the 1st and 2nd waiters only.
If we were to wake the 3rd, 5th and 7th waiters also, then the queue
would reduce in length very quickly, if we assume generally uniform
service times. (If the head of the queue is X, then we wake only that
one process and I'm not proposing we change that). That would mean queue
jumping right? Well thats what already happens in other circumstances,
so there cannot be anything intrinsically wrong with allowing it, the
only question is: would it help?
We need not wake the whole queue, there may be some generally more
beneficial heuristic. The reason for considering this is not to speed up
Shared requests but to reduce the queue length and thus the waiting time
for the Xclusive requestors. Each time a Shared request is dequeued, we
effectively re-enable queue jumping, so a Shared request arriving during
that point will actually jump ahead of Shared requests that were unlucky
enough to arrive while an Exclusive lock was held. Worse than that, the
new incoming Shared requests exacerbate the starvation, so the more
non-adjacent groups of Shared lock requests there are in the queue, the
worse the starvation of the exclusive requestors becomes. We are
effectively randomly starving some shared locks as well as exclusive
locks in the current scheme, based upon the state of the lock when they
make their request. The situation is worst when the lock is heavily
contended and the workload has a 50/50 mix of shared/exclusive requests,
e.g. serializable transactions or transactions with lots of
Goal: Reduce the total time that CommitTransaction() waits for
5. Reduce the time that GetSnapshotData holds ProcArray lock. To do
this, we split the ProcArrayLock into multiple partitions (as suggested
by Alvaro). There are comments in GetNewTransactionId() about having one
spinlock per ProcArray entry. This would be too many and we could reduce
contention by having one lock for each N ProcArray entries. Since we
don't see too much contention with 100 users (default) it would seem
sensible to make N ~ 120. Striped or contiguous? If we stripe the lock
partitions then we will need multiple partitions however many users we
have connected, whereas using contiguous ranges would allow one lock for
low numbers of users and yet enough locks for higher numbers of users.
6. Reduce the number of times ProcArrayLock is called in Exclusive mode.
To do this, optimise group commit so that all of the actions for
multiple transactions are executed together: flushing WAL, updating CLOG
and updating ProcArray, whenever it is appropriate to do so. There's no
point in having a group commit facility that optimises just one of those
contention points when all 3 need to be considered. That needs to be
done as part of a general overhaul of group commit. This would include
making TransactionLogMultiUpdate() take CLogControlLock once for each
page that it needs to access, which would also reduce contention from
(1) and (2) can be patched fairly easily for 8.3. I have a prototype
patch for (1) on the shelf already from 6 months ago.
(3), (4) and (5) seem like changes that would require significant
testing time to ensure we did it correctly, even though the patches
might be fairly small. I'm thinking this is probably an 8.4 change, but
I can get test versions out fairly quickly I think.
(6) seems definitely an 8.4 change.
pgsql-hackers by date
|Next:||From: Devrim GÜNDÜZ||Date: 2007-07-30 19:23:23|
|Subject: Re: Machine available for community use|
|Previous:||From: Bruce Momjian||Date: 2007-07-30 12:47:46|
|Subject: Re: default_text_search_config and expression