Re: How a sort works?
- From: "Finn Jorgensen" <finn.oracledba@xxxxxxxxx>
- To: DEEDSD@xxxxxxxxxxxxxx
- Date: Sat, 1 Dec 2007 17:16:29 -0500
What I often times do when I run into this is I wrap the statement including
the order by in a select count(*) from () like this :
select count(*) from (
select... original SQL
order by ...
)
and make sure the explain plan is the same. Whether that statement is fast
and slow, you'll know if the problem is the sort or the data return to the
client. Then you can act accordingly, or at least tell the customer what's
going on.
Finn
On Nov 30, 2007 5:45 PM, <DEEDSD@xxxxxxxxxxxxxx> wrote:
>
> First, allow me to apologize for our Lotus Notes throwing my ID in all
> caps. If someone knows how to turn that annoyance off, please tell me!
>
> We have a fairly sizeable sort order-by (1.2 million rows).
> V$session_longops shows that it takes a massive amount of time in this
> step. I have traced the session with 10046. I am working on tracing this
> with 10032 and 10033 events turned on, as well as statistics_level=all so I
> can get timings in the stat lines.
>
> I see a big long fetch, then a bunch (190) repetitions of more data to
> client of around 2,000 bytes (all but one are of short duration), with some
> direct path reads sprinkled in. I see a more data to client of a very long
> duration (in the middle of the 190), more of short duration, and then
> another long fetch (these steps repeat).
>
>
> It looks like the really long "more data to client" time is suspiciously
> close to the fetch e - c time.
>
> The way I'm thinking is this: Oracle sorts some of the data, building a
> partial resultset, spews a bunch or rows off to the client, and waits for
> some sort output buffer to clear before throwing more rows into it. The c
> time in the fetch is primarily sort time, and v$session longops is
> attributing all the time to the sort, in a rather misleading fashion. The
> bottleneck is really how much data the client can take. Am I correct in
> this line of thinking? I'm also not sure what the long-duration more data
> to client is telling me.
>
> Solaris 5.10, Oracle 9.2.0.8
>
> <snip>
> WAIT #3: nam='SQL*Net more data to client' ela= 19 p1=1413697536 p2=2000
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 84 p1=1413697536 p2=2001
> p3=0
> WAIT #3: nam='direct path read' ela= 104 p1=203 p2=16661 p3=1
> WAIT #3: nam='SQL*Net more data to client' ela= 148379 p1=1413697536
> p2=2001 p3=
> 0
> WAIT #3: nam='SQL*Net more data to client' ela= 22 p1=1413697536 p2=2003
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 33 p1=1413697536 p2=2005
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 38 p1=1413697536 p2=1997
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 35 p1=1413697536 p2=2004
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 36 p1=1413697536 p2=1996
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 31 p1=1413697536 p2=2003
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 32 p1=1413697536 p2=2002
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 28 p1=1413697536 p2=1998
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 28 p1=1413697536 p2=2001
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 36 p1=1413697536 p2=2001
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 29 p1=1413697536 p2=2005
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 30 p1=1413697536 p2=1997
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 29 p1=1413697536 p2=2002
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 45 p1=1413697536 p2=2000
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 30 p1=1413697536 p2=2001
> p3=0
> WAIT #3: nam='SQL*Net more data to client' ela= 37 p1=1413697536 p2=2006
> p3=0
> FETCH
> #3:c=80000,e=223197,p=48,cr=0,cu=0,mis=0,r=10000,dep=0,og=3,tim=3397704949
> 096
> *** 2007-11-27 14:27:24.028
> WAIT #3: nam='SQL*Net message from client' ela= 55007450 p1=1413697536
> p2=1 p3=0
>
>
Other related posts:
