RE: impdp performance in 11.1 vs 11.2
Date: Wed, 20 Aug 2014 20:44:22 -0400
Message-ID: <062a01cfbcd9$0de5c1c0$29b14540$_at_rsiz.com>
I believe you said the same thing. One thread of index creation, but “each index in parallel up to the degree of parallelism of the job.” The way I read that is serially doing one index at a time at the parallelism of PX slaves matching the parallelism of the job.
When under time pressure you can work around this by NOT having impdp do the index creates. Arrange to have your biggest few tables loaded first and run the index creation scripts in multiple threads table by table so you get some drafting of the read part of index build because it has just been done into cache. 11.2 plus this *MAY* mean a special table read job to avoid adaptive direct read so the table does pump through cache and you avoid multiple repetitive block cleanouts and multiple physical reads.
IF you have plenty of time, avoid this slippery slope toward CTD. (tm gaja).
From: oracle-l-bounce_at_freelists.org [mailto:oracle-l-bounce_at_freelists.org] On Behalf Of Seth Miller
Sent: Wednesday, August 20, 2014 6:55 PM
To: rjamya_at_gmail.com
Cc: mschmitt_at_uchicago.edu; Oracle-List; Andy Wattenhofer
Subject: Re: impdp performance in 11.1 vs 11.2
Raj,
You may be thinking of an older version. This is from an 11.2 Oracle white paper.
During export, when there are two or more worker processes, data and metadata unloading proceed in parallel. Data Pump will also automatically build each index in parallel up to the degree of parallelism of the job. The permanent parallel degree of the index itself remains unchanged from its source value.
Seth Miller
On Wednesday, August 20, 2014, rjamya <rjamya_at_gmail.com> wrote:
Index creation never runs in parallel worker processes, only data load can run in parallel. Even metasata load/unload is run by a single worker process. You will need to trace impdp to find why something is fast and slow. Google for 'expdp trace'.
I think even on 12.1 index creation is run by only one thread. It may use PX slaves if index was originally parallel by only one worker thread does that whilest others are idle.
Raj
-- http://www.freelists.org/webpage/oracle-lReceived on Thu Aug 21 2014 - 02:44:22 CEST