Обсуждение: Checkpoint question
I understand checkpoint code doing something like this: Get RedoRecPtr;Flush all dirty buffers no matter what's its LSN;Write down checkpoint xlog record; So I wonder is it possible flush only dirty buffers with LSN < RedoRecPtr to improve checkpoint caused delay? Because even we flush every dirty buffers, we still have to replay from the RedoRecPtr. Of course, this only applies to non-critical checkpoints (critical ones like startup and shutdown). Regards, Qingqing
On Wed, 2006-01-11 at 18:24 -0500, Qingqing Zhou wrote: > I understand checkpoint code doing something like this: > > Get RedoRecPtr; > Flush all dirty buffers no matter what's its LSN; > Write down checkpoint xlog record; > > So I wonder is it possible flush only dirty buffers with LSN < RedoRecPtr > to improve checkpoint caused delay? Because even we flush every dirty > buffers, we still have to replay from the RedoRecPtr. Of course, this only > applies to non-critical checkpoints (critical ones like startup and > shutdown). Probably good idea to read Gray & Reuter or Vekum & Vossen books on transactional systems theory before any such discussion. Incidentally, it was suggested to me that we write odd/even numbered blocks on alternate checkpoints as a way of reducing checkpoint impact. Apparently this has been implemented on another RDBMS in a galaxy far, far away. But I have enough to do right now. Best Regards, Simon Riggs
On Wed, 11 Jan 2006, Simon Riggs wrote: > > Probably good idea to read Gray & Reuter or Vekum & Vossen books on > transactional systems theory before any such discussion. > So can you give me some hints why my thoughts are just wrong? Regards, Qingqing
Qingqing Zhou <zhouqq@cs.toronto.edu> writes:
> So I wonder is it possible flush only dirty buffers with LSN < RedoRecPtr
> to improve checkpoint caused delay?
Certainly not. If LSN > RedoRecPtr then you know the buffer contains
some changes more recent than the checkpoint, but you cannot tell
whether it also contains changes older than the checkpoint. For
correctness you must flush it.
It'd be possible to do something like this: after establishing
RedoRecPtr, make one quick pass through the buffers and make a list of
what needs to be dumped at that instant. Then go back and do the actual
I/O for only those buffers. I'm dubious that this will really improve
matters though, as the net effect is just to postpone I/O that will
happen anyway soon after the checkpoint (as soon as the bgwriter goes
back to normal activity).
regards, tom lane
On Wed, 11 Jan 2006, Tom Lane wrote: > Qingqing Zhou <zhouqq@cs.toronto.edu> writes: > > So I wonder is it possible flush only dirty buffers with LSN < RedoRecPtr > > to improve checkpoint caused delay? > > Certainly not. If LSN > RedoRecPtr then you know the buffer contains > some changes more recent than the checkpoint, but you cannot tell > whether it also contains changes older than the checkpoint. For > correctness you must flush it. > Right. > It'd be possible to do something like this: after establishing > RedoRecPtr, make one quick pass through the buffers and make a list of > what needs to be dumped at that instant. Then go back and do the actual > I/O for only those buffers. I'm dubious that this will really improve > matters though, as the net effect is just to postpone I/O that will > happen anyway soon after the checkpoint (as soon as the bgwriter goes > back to normal activity). > Looks like a good idea. I don't worry too much about the problem you mentioned. AFAIK, checkpoint has two targets: (1) cleanup buffer pool; (2) reduce recovery time; For (2), it is clear that the above idea will work since the recovery will always read the data page to check its LSN -- the is the source of the cost. For (1), we have bgwriter, and part of reason it is desiged is to cleanup buffer pool. On a separate matter, it will be interesting to add a io-status-dector in bgwriter to make it know when the disk is free so a cleanup is valuable. Regards, Qingqing
On Wed, 11 Jan 2006, Tom Lane wrote: > > It'd be possible to do something like this: after establishing > RedoRecPtr, make one quick pass through the buffers and make a list of > what needs to be dumped at that instant. Then go back and do the actual > I/O for only those buffers. I'm dubious that this will really improve > matters though, as the net effect is just to postpone I/O that will > happen anyway soon after the checkpoint (as soon as the bgwriter goes > back to normal activity). > We could extend the grammar to CHECKPOINT [FULL/PARTIAL] to let the user decide what to do. If PARTIAL, an important reason is for fast recovery. Regards, Qingqing
On Wed, 2006-01-11 at 20:46 -0500, Qingqing Zhou wrote: > > On Wed, 11 Jan 2006, Simon Riggs wrote: > > > > > Probably good idea to read Gray & Reuter or Vekum & Vossen books on > > transactional systems theory before any such discussion. > > > So can you give me some hints why my thoughts are just wrong? Your thoughts are very often good ones, ISTM. This area is a minefield of misunderstanding, as are many others, but in this case some external wisdom is readily available. Those books helped me understand things better; it was wrong of me to assume you hadn't already read them. Best Regards, Simon Riggs
On Wed, 2006-01-11 at 22:33 -0500, Qingqing Zhou wrote: > On Wed, 11 Jan 2006, Tom Lane wrote: > > It'd be possible to do something like this: after establishing > > RedoRecPtr, make one quick pass through the buffers and make a list of > > what needs to be dumped at that instant. Then go back and do the actual > > I/O for only those buffers. I'm dubious that this will really improve > > matters though, as the net effect is just to postpone I/O that will > > happen anyway soon after the checkpoint (as soon as the bgwriter goes > > back to normal activity). > > > Looks like a good idea. I don't worry too much about the problem you > mentioned. AFAIK, checkpoint has two targets: (1) cleanup buffer pool; (2) > reduce recovery time; I think its a good idea, but agree it does not save much in practice. The only buffers this will miss are ones that were clean throughout the whole of the last checkpoint cycle, yet have been dirtied between the start of the checkpoint pass and when the pass reaches it. Given the relative durations of those two intervals, I would guess that this would yield very few buffers. Further, if you miss a buffer on one checkpoint it will not be able to avoid being written at the next. If we write the buffer again in next checkpoint cycle then we combine the two I/Os and save effort. If the buffer is not written to in the next cycle, and this seems likely since it wasn't written to in the last, we do not avoid I/O, we just defer it to next checkpoint. So the only buffer I/O we would save is for buffers that - are not written to in checkpoint cycle, n (by definition) - are written to *during* the checkpoint - are written to again during the next checkpoint cycle, n+1 You could do math, or measure that, though my guess is that this wouldn't save more than a few percentage points on the checkpoint process. To compile the list, you'd need to stop all buffer write activity while you compile it, which sounds a high price for the benefit. > For (2), it is clear that the above idea will work since the recovery will > always read the data page to check its LSN -- the is the source of the > cost. For (1), we have bgwriter, and part of reason it is desiged is to > cleanup buffer pool. Deferring I/O gains us nothing in the long run, though would speed up recovery time by a fraction - but then crash recovery time is not much an issue is it? If it is, there are other optimizations. Best Regards, Simon Riggs
Simon Riggs <simon@2ndquadrant.com> writes:
>> On Wed, 11 Jan 2006, Tom Lane wrote:
>>> It'd be possible to do something like this: after establishing
>>> RedoRecPtr, make one quick pass through the buffers and make a list of
>>> what needs to be dumped at that instant. Then go back and do the actual
>>> I/O for only those buffers.
> To compile the list, you'd need to stop all buffer write activity while
> you compile it, which sounds a high price for the benefit.
Not really --- I was only thinking of narrowing the window for "extra"
writes to get in, not removing the window entirely. Don't need any sort
of global lock for that.
But I agree with your analysis that the extra cycles won't save much in
practice. The objection I see is that two lock cycles on each targeted
buffer are a nontrivial expense in SMP machines.
regards, tom lane
On Thu, 12 Jan 2006, Simon Riggs wrote: > > The only buffers this will miss are ones that were clean throughout the > whole of the last checkpoint cycle, yet have been dirtied between the > start of the checkpoint pass and when the pass reaches it. I agree on the analysis but I am not sure current interval of doing a checkpoint. So it depends. If the checkpoint on an io-intensive machine the interval I guess would not be small. Also, in this environment, one more round of lock cycle should be relatively cheap. But we currently don't have any numbers on hand ... Regards, Qingqing
On Thu, Jan 12, 2006 at 04:50:30AM -0500, Qingqing Zhou wrote: > > > On Thu, 12 Jan 2006, Simon Riggs wrote: > > > > The only buffers this will miss are ones that were clean throughout the > > whole of the last checkpoint cycle, yet have been dirtied between the > > start of the checkpoint pass and when the pass reaches it. > > I agree on the analysis but I am not sure current interval of doing a > checkpoint. So it depends. If the checkpoint on an io-intensive machine > the interval I guess would not be small. Also, in this environment, one > more round of lock cycle should be relatively cheap. But we currently > don't have any numbers on hand ... It sounds like worrying about this would be much more interesting on a machine that is seeing both a fairly heavy IO load (meaning checkpoint will both take longer and affect other workloads more) and is seeing a pretty high rate of buffer updates (meaning that we'd likely do a bunch of extra work as part of the checkpoint if we didn't take note of exactly what buffers needed to be flushed). Unfortunately I don't think there's any way for the backend to know much about either condition right now, so it couldn't decide when it made sense to make a list of buffers to flush. Maybe in the future... As for the questionable benefit to delaying work for bgwriter or next checkpoint, I think there's a number of scenarios where it would make sense. A simple example is doing some kind of processing once a minute that's IO intensive with default checkpoint timing. Sometimes a checkpoint will occur at the same time as the once-a-minute process, and in those cases reducing the amount of work the checkpoint does will definately help even out the load on the machine. -- Jim C. Nasby, Sr. Engineering Consultant jnasby@pervasive.com Pervasive Software http://pervasive.com work: 512-231-6117 vcard: http://jim.nasby.net/pervasive.vcf cell: 512-569-9461
On Thu, 12 Jan 2006, Jim C. Nasby wrote: > > It sounds like worrying about this would be much more interesting on a > machine that is seeing both a fairly heavy IO load (meaning checkpoint > will both take longer and affect other workloads more) and is seeing a > pretty high rate of buffer updates (meaning that we'd likely do a bunch > of extra work as part of the checkpoint if we didn't take note of > exactly what buffers needed to be flushed). Unfortunately I don't think > there's any way for the backend to know much about either condition > right now, so it couldn't decide when it made sense to make a list of > buffers to flush. Maybe in the future... > The senario you mentioned is happened in many OLTP applications. No need for backend to know this -- we can leave the decision to the DBA: CHECKPOINT FULL or CHECPOINT PARTIAL. If you got some machines can observe its CHECKPOINT duration, that would be sweet. Regards, Qingqing
On Thu, Jan 12, 2006 at 05:00:49PM -0500, Qingqing Zhou wrote: > > > On Thu, 12 Jan 2006, Jim C. Nasby wrote: > > > > > It sounds like worrying about this would be much more interesting on a > > machine that is seeing both a fairly heavy IO load (meaning checkpoint > > will both take longer and affect other workloads more) and is seeing a > > pretty high rate of buffer updates (meaning that we'd likely do a bunch > > of extra work as part of the checkpoint if we didn't take note of > > exactly what buffers needed to be flushed). Unfortunately I don't think > > there's any way for the backend to know much about either condition > > right now, so it couldn't decide when it made sense to make a list of > > buffers to flush. Maybe in the future... > > > > The senario you mentioned is happened in many OLTP applications. No need > for backend to know this -- we can leave the decision to the DBA: > CHECKPOINT FULL or CHECPOINT PARTIAL. If you got some machines can observe > its CHECKPOINT duration, that would be sweet. Maybe I'm missing something here, but wouldn't that only help if you were manually issuing checkpoints? -- Jim C. Nasby, Sr. Engineering Consultant jnasby@pervasive.com Pervasive Software http://pervasive.com work: 512-231-6117 vcard: http://jim.nasby.net/pervasive.vcf cell: 512-569-9461