Parallel Index creation takes more time...!!

Question

OS - Windows 2008 Server R2
Oracle - 10.2.0.3.0
My table size is - 400gb
Number of records - 657,45,95,123
my column definition first_col varchar2(22) ; -> I am creating index on this column
first_col -> actual average size of column value is 10
I started to create index on this column by following command
CREATE INDEX CALL_GROUP1_ANO ON CALL_GROUP1(A_NO) LOCAL PARALLEL 8 NOLOGGING COMPRESS ;
-> In my first attempt after three hours I got an error :
ORA-01652: unable to extend temp segment by 128 in tablespace TEMP
So I increased the size of temp tablespace to 380GB ,Because i expect the size of first_col index this much.
-> In my second attempt Index creation is keep going even after 17 hours...!!
Now the usage of temp space is 162 GB ... still it is growing..
-> I checked EM Advisor Central ADDM :
it says - The PGA was inadequately sized, causing additional I/O to temporary tablespaces to consume significant database time.
1. why this takes this much of Temp space..?
2. It is required this much of time to CREATE INDEX in parallel processing...? more than 17 hrs
3. How to calculate and set the size of PGA..?

Answer 1

OraFighter wrote:
Oracle - 10.2.0.3.0
My table size is - 400gb
Number of records - 657,45,95,123
my column definition first_col varchar2(22) ; -> I am creating index on this column
first_col -> actual average size of column value is 10
I started to create index on this column by following command
CREATE INDEX CALL_GROUP1_ANO ON CALL_GROUP1(A_NO) LOCAL PARALLEL 8 NOLOGGING COMPRESS ;
Now the usage of temp space is 162 GB ... still it is growing..The entire data set has to be sorted - and the space needed doesn't really vary with degree of parallelism.
6,574,595,123 index entries with a key size of 10 bytes each (assuming that in your choice of character set one character = one byte) requires per row approximately
4 bytes row overhead 10 bytes data, 2 bytes column overhead for data, 6 bytes rowid, 2 bytes column overhead for rowid = 24 bytes.
For the sorting overheads, using the version 2 sort, you need approximately 1 pointer per row, which is 8 bytes (I assumed you're on 64 bit Oracle on this platform) - giving a total of 32 bytes per row.
32 * 6,574,595,123 / 1073741824 = 196 GB
You haven't said how many partitions you have, but you might want to consider creating the index unusable, then issuing a rebuild command on each partition in turn. From "Practical Oracle 8i":
<blockquote>
In the absence of partitioned tables, what would you do if you needed to create a new index on a massive data set to address a new user requirement? Can you imagine the time it would take to create an index on a 450M row table, not to mention the amount of space needed in the temporary segment. It's the sort of job that you schedule for Christmas or Easter and buy a couple of extra discs to add to the temporary tablespace.
With suitably partitioned tables, and perhaps a suitably friendly application, the scale of the problems isn't really that great, because you can build the index on each partition in turn. This trick depends on a little SQL feature that appears to be legal even though I haven't managed to find it in the SQL reference manual:
     create index big_new_index on partitioned_table (colX)
     local
     UNUSABLE
     tablespace scratchpad
The key word is UNUSABLE. Although the manual states that you can 'alter' an index to be unusable, it does not suggest that you can create it as initially unusable, nevertheless this statement works. The effect is to put the definition of the index into the data dictionary, and allocate all the necessary segments and partitions for the index - but it does not do any of the real work that would normally be involved in building an index on 450M rows.
</blockquote>
(The trick was eventually documented a couple of years after I wrote the book.)
Regards
Jonathan Lewis