Subject | Re: commit after insert is very slow (bis - new thread) |
---|---|
Author | svanderclock |
Post date | 2010-03-04T20:29:33Z |
dear ann !
i understand, and you was right, because on the task manager i see with a low page buffer than the fbserver.exe process read / write during the execution time !
now i encountered another strange behavior, it's when i setup a big page buffer
page buffer 2048: execute time 156ms
page buffer 20480: execute time 156ms
page buffer 100000: execute time 15288 ms !!!
any idea why when i setup a big page buffer then the execute time go down by 100 ! the commit time don't change..
many thanks by advance for your precious help !!
stephane
i understand, and you was right, because on the task manager i see with a low page buffer than the fbserver.exe process read / write during the execution time !
now i encountered another strange behavior, it's when i setup a big page buffer
page buffer 2048: execute time 156ms
page buffer 20480: execute time 156ms
page buffer 100000: execute time 15288 ms !!!
any idea why when i setup a big page buffer then the execute time go down by 100 ! the commit time don't change..
many thanks by advance for your precious help !!
stephane
--- In firebird-support@yahoogroups.com, "Ann W. Harrison" <aharrison@...> wrote:
>
> svanderclock wrote:
> >
>
> Reworked your data as a chart. For 40 inserts and 40 deletes:
>
> Page Size Page Buffers Execute Time Commit Time
>
> 4096 75 9252ms 78ms
> 4096 2048 156ms 2543ms
> 16384 2048 172ms 1529ms
> >
> >
> > why with a lower page buffer the commit time is very fast but the execution time is very slow and with a more larger buffer it's the opposite ?
>
> With 75 buffers, you were writing out most of the pages before
> the commit. At most, you wrote 75 pages on commit. With 2048
> pages, you're writing that much more on commit. When you go
> to a large page size, you write 1/4 as many pages, each of which
> is 4 times larger. Bigger writes are slightly more expensive,
> but fewer larger writes seems help on commit. In all cases,
> probably got several upper level index pages that don't change,
> so the actual cost per page written on commit depends on page
> size.
>
> The depth of the index matters much less for writes - finding
> the location is relatively cheap compared with writing back
> a page. Index depth shows up as a problem mostly on reads.
>
>
> Good luck,
>
> Ann
>