Single block read for Sort

Question

I’m creating a temp table with following SQL. Table temp2 is 50GB in size with 120M rows. I noticed during SORT operation , it’s performing lots of single block reads on Temp tablespace. Why it’s doing single block reads on temp??
Also, Is there any other way to make this quicker?
I'm on 10gR2 with Auto PGA/SGA. PGA Aggregate is 5GB and SGA is 12GB.
create table t_temp nologging parallel(degree 4 )
as
select /*+ parallel(t1,4) */ * from temp2 partition(P2008)  t1
order by custkey
    SID Wait State    EVENT                               P1                         P2              P3                    SEQ# % TotTime TEventTime(ms)  DistEvnts Avgtime(ms)/Evnt
   2165 WAITING       direct path read temp               file number= 5001          first dba=      block cnt= 1                     .02           .253          1             .253
                                                                                     88791
   2165 WAITING       direct path read temp               file number= 5001          first dba=      block cnt= 1                     .02           .253          1             .253
                                                                                     412771
   2165 WAITING       direct path read temp               file number= 5001          first dba=      block cnt= 1                     .02           .253          1             .253
                                                                                     421465
   2165 WAITING       direct path read temp               file number= 5001          first dba=      block cnt= 1                     .02           .253          1             .253
                                                                                     691141
   2165 WAITING       direct path read temp               file number= 5001          first dba=      block cnt= 1                     .02           .253          1             .253
                                                                                     1295425
Here is  Temp table space properties…
Name                    INITK    NEXTK MINEX MAXEX PCTI CNTS      EXMGMT     ALOCTYPE  SEGMGM STATUS
TEMP                     1024     1024     1          0 TEMPORARY LOCAL      UNIFORM   MANUAL ONLINEEdited by: user4529833 on Feb 5, 2009 9:05 AM

Answer 1

user4529833 wrote:
Thanks Jonathan.
I will do that and post the results... One questions though, how to correct this unexpected side-effect of asynchronous reads... ( Assuming it's the case )
If the assumption is correct, then it's not really something that needs to be corrected. Because of the overlapping of I/Os, there may be lots of I/Os that never make it into wait time, so a single process could be going:
give me 12 blocks
give me 12 blocks
give me 12 blocks
give me 12 blocks
give me a block
of which you see only the "give me a block".
That's why the 10033 is an important diagnostic - it will help you decide whether or not "everything" is a single block read, or whether the single block reads are a very tiny fraction of the total I/O that also happens to become visible for some reason.
how does oracle perform sort for SQLs I have given..
1) read & sort the data for custid coulmn with rowids and dump it temp (quite likely)
2) read the rowids from temp and start loading row to temp table..I'd have to check the execution to give you a definite answer but the basic strategy would be:
<ul>
Two sets of PX slaves.
Set one reads a chunk of the source table picking up all the relevant columns for the result table.
Because of the ORDER BY clause layer one does a range-based distribution of data to the second set of slaves - along the lines of "slave 1 gets custid less than 1M", "slave 2 gets custid between 1M and 2M" and so on. Because entire rows are passed between the two sets of slaves, you can see a LOT of PQ traffic going on. (People sometimes increase the "PX message size" parameter because of the volume of traffic and the size of the rows).
Set two slaves accumulate, and sort the incoming data sets. Each sorted row carries ALL the columns, there is no "sort the custkey/rowids only" strategy. The sorting may spill to temp - given your "select *", the amount of temp space used may actually be greater than the size of the source table.
Each set two slave creates its own "private table" with its results.
The query coordinator links the "private table" metadata in the data dictionay in the right order to create the final table.
</ul>
Regards
Jonathan Lewis
http://jonathanlewis.wordpress.com
http://www.jlcomp.demon.co.uk
"Science is more than a body of knowledge; it is a way of thinking"
Carl Sagan
To post code, statspack/AWR report, execution plans or trace files, start and end the section with the tag {noformat}{noformat} (lowercase, curly brackets, no spaces) so that the text appears in fixed format.