Comments on: Upscaling Your Database Application Performance: The Array Interface

By: Dominic Brooks

Dominic Brooks — Fri, 02 Oct 2009 15:44:44 +0000

@Chris In my comment above, I linked to an example done with the clumsy struct / array descriptor classes (clumsy in my non-existent java anyway).Here's a link to that example - it was twice as fast as the timings for my batch size of 50.

By: Chris

Chris — Wed, 23 Sep 2009 21:08:27 +0000

I’ve looked at the attached Java code and this is using the JDBC batch API available with JDBC 2.0 onwards. I thought this might have been the native array interface, this is faster as it allows arrays of values to be passed to stored procedures, which can then use bulk binds. Unsurprisingly there is a good example on Ask Tom: http://asktom.oracle.com/pls/asktom/f?p=100:11:0::::P11_QUESTION_ID:712625135727. Both techniques eliminate network round trips, but the second will be a lot more cursor friendly. The obvious advantage of the JDBC 2.0 batch API is that it is portable and most ORMs (object relational mapping) tools and frameworks support this. As far as I’m aware use native array interface requires hand cranked JDBC code.

By: Chris

Chris — Mon, 24 Aug 2009 20:51:42 +0000

@Greg Rahn I believe that the batching API available with JDBC 2.0 onwards 'batches' DML statements by putting a BEGIN and END around them, hence submitting them to the database in an anonymous PL/SQL block. IMHO techniques to reduce network round trips, which the JDBC batching API and array interface help to achieve, should be added to the scalability and performance "Holy trinity" of connection management, cursor management and good schema design.

By: Greg Rahn

Greg Rahn — Thu, 03 Jul 2008 04:53:41 +0000

@oracle tom

I am not familiar with the native JDBC 2.0 implementation method, so I would suggest that you try both and see how they compare. Just turn sql trace on in your test app and then tkprof the trace file.

By: oracle tom

oracle tom — Mon, 30 Jun 2008 10:57:55 +0000

Is there a difference in performance if the JDBC 2.0 batching or the oracle implemntation is used? (in java)

By: jan van mourik

jan van mourik — Sat, 27 Oct 2007 19:32:15 +0000

Ran a few more tests, this time connecting to the database on my laptop (10g). Also, I added a table EMP2, with a FK from emp2.mgr to emp.empno. So I’m using two preparedStatements. This to see if the batching would cause any problems with the FK. The rows inserted look the same otherwise, and I’m inserting the same number of rows in both tables. The results:

10,000 rows into both emp and emp2
batch elapsed in sec
1 3.266 sec
10 0.563
50 0.328
100 0.297
200 0.453
1000 0.437

100,000 rows
1 40.9 sec
100 3.344

1,000,000 rows
1 376.532 sec
100 45
200 36.922
500 32.875
1000 29.563

Big difference from my previous test, where I accessed a database on another machine over a very slow network. Nice…

jan

By: jan van mourik

jan van mourik — Fri, 26 Oct 2007 19:21:28 +0000

Thanks Greg. It was also part laziness on my side. Since I’m not a java programmer I had some trouble getting it to work…
Btw, my results inserting 10,000 rows (emp + pk, no fk):
Batch size Time(s)
1 652
10 65
20 32
40 16
50 14
100 7
200 3.8
400 3.47
500 2.89
1000 2.7

Notice how it scales up to batch size 200!!

Thanks Greg!!!

jan

By: Greg Rahn

Greg Rahn — Thu, 25 Oct 2007 18:00:52 +0000

I don't think I have the trace files any more but the code is there and one can easily run the test case. The batchInsert.java has the sql_trace line in it.

By: jan van mourik

jan van mourik — Thu, 25 Oct 2007 17:35:54 +0000

Nice, very interesting!
I’d like to see the tkprof output for a few testcases too (or Hotsos profiler if possible). Did you trace these testcases?

Regards, jan

By: thangadurai

thangadurai — Wed, 26 Sep 2007 07:33:43 +0000

its very userful to all, who are working in both 10g And 11g databases.