Re: Extended checkpoint times

Peter Tashkoff wrote: > I had a problem in November of checkpoints blowing out to 10 minutes or so > during the evening like your manifestation. > I was advised from tech support that checkpoints can blow out two ways. > 1. When checkpoint is requested, Online freezes all sessions for users not > in a critical section and waits for remainder to exit critical section > before freezing their sessions. > A critical section is most likely to be a physical write to disk. > > 2. If Online is carrying out some large unusual process, like a large = > table load, this can also cause checkpoints to blow out. > It is possible that a disk problem could be the cause. To investigate > this possibility, during an interval when system is frozen do the > following; > 1. ID the sessions in critical sections. > 2. onstat -g sessid to ID the tables being connected to. > 3, ID the dbspace of the tables concerned. > 4. ID the raw partition that this relates to. > In our case the sysadmin had been messing around with disks getting the > machine ready for a solaris upgrade. I don't know why this might have > affected online but the problem did go away after the disk got sorted out. > HTH > Peter Tashkoff <tashkop@iname.com> > >>> anonymous masquerading as Mark said >>> > Hi, > We are trying to investigate a series of overly long checkpoints on one > of our informix databases, here is > a snippet of the online.log file detailing the events: > 21:16:55 Checkpoint Completed: duration was 76 seconds. > 21:22:04 Checkpoint Completed: duration was 6 seconds. > 21:27:16 Checkpoint Completed: duration was 8 seconds. > 21:40:15 Checkpoint Completed: duration was 476 seconds. > 21:45:29 Checkpoint Completed: duration was 10 seconds. > 21:50:36 Checkpoint Completed: duration was 2 seconds. > 21:55:42 Checkpoint Completed: duration was 3 seconds. > 22:03:22 Checkpoint Completed: duration was 155 seconds. We see this exact situation during an archive. Occassionally during an archive a checkpoint will last from 100->800 seconds while our normal checkpoint duration is 5-7 seconds. Once this hung forever which turned out to be a deadlock between two threads each of which had acquired a latch on one of the Physical Log buffers and wanted a latch on the other which of course was already help by the other thread this caused a deadlock which was not detected and resolved between two threads in critical sections stalling the checkpoint indefinitely. I have reported both of these situations but there is as yet no resolution nor a comprehensive explanation of what is happening in the case that you describe where the checkpoint eventually completes. Art S. Kagel