Wednesday, July 12, 2017

Profiling a Java + JDBC Application

NetBeans


First, there's NO Java coding needed nor Java source code needed to profile a Java program this way.  NetBeans added this a while back up I just found it recently.  The ability to attach to any Java program and profile the SQL going across JDBC. The dev team's blog on it is here: http://jj-blogger.blogspot.nl/2016/05/netbeans-sql-profiler-take-it-for-spin.html


SQLcl

SQLcl is our Java library for scripting sql/sql scripts that has been in SQLDev since day 0 back in '05/06.  We factored that and wrapped a cmd line around it.  This makes it easier to test for features, regressions, performance,....  as well to give a new cmd line with extended features.  This library is also what performs the Grid Infra installs these days as well embedded in Oracle REST Data Services. It's quite proven and tested. This is all Java bases using plain JDBC to talk to the database.  It's no different than any java based application which means anything done to profile it is applicable to any java program like say sqldev , ords, custom jdbc , any java program.




Profiling


This new feature in Netbeans is very simple to use and there's no need to have the sources of the jsvs code.  Off the Profile menu - > Attach to External Process






Then set the Profile to SQL Queries


Click Attach, which shows a list of running java processes.  This is what SQLcl will look like.





Running the Program

Now once JDBC traffic starts being issued, it's captured with timings and occurrences of that statement along with the Java stack for where the call originated. Next up is the hardest part, what the heck does all this data mean? When is fast , fast enough?

What to change?


Below is what an APEX install looks like on my laptop during the middle of the process.  There's a lot of data to look at. The slowest statement is the dbms registry validation. Is that bad, can it be sped up? Probably not. The most called is the check for DBMS_OUPUT. Can that be reduced? Also, probably not.

This is when knowledge of the code and intended actions are critical.  For me, getting SQLcl the program from 19m down to 7m was fast enough. That was done with zero changes to the APEX install scripts but just from watching the traffic going to the database and analyzing that.

Change #1 : SQLcl was name resolving every create or replace <PLSQL OBJECT> then checking for errors on that object.  Much faster is to simply check count(1) from user_errors without the name resolution.  When there's no errors in the user_errors table, there's no need to name resolve. So that entire path was shortened.  It's visible in this stack with the 1,106 time "select count(1) cnt from user_errors" was called.  

Change #2: DBMS_OUTPUT was being called after any/all commands to the database. That was reduced to only calls that could invoke some output. For Example, alter session doesn't need to be checked. That change reduced the number of db calls being issued at all. Fastest call is the ones you don't make.

and on and on.

Nothing is more important than the knowledge of the intended outcome.








Tuesday, July 11, 2017

SQLcl 17.2

New Versioning Scheme

Starting with this release the numbering scheme is changed.  All releases will now be the YEAR<period>Quarter<period>build numbers.

So the new SQLcl is 17.2.0.184.0917.  

Breaking that down. 
  • 17   - Year
  • 2     - Quarter
  • 0     -  Patch number
  • 184 - Day in Julian
  • 0917 - hour and minute the build was done.

New Features

Securing Literals  which was introduced here : http://krisrice.blogspot.com/2015/09/sqlcl-more-secure-now-with-rest.html so this is not new.  What is new is controls over when it's done.  It was set so that SQLcl did secure all literals for anything that was issued. Now there's a control for when/how deep to check.

The default is any anonymous block less than 10 lines will be scrubbed automatically.  This will catch the majority of uses.  To ratchet up what is checked  "set secureliterals ON" will secure every block completely.  There is a performance impact to this if there are very large block such as in the APEX installation which has some blocks over 1k in size.

The opposite is there also to disable this feature: set secureliterals OFF


Here's an example of what happens. The 'abcxyz' is removed and turned into a bind :SqlDevBind1ZInit1


SQL> declare
  2    l_local varchar2(20);
  3  begin
  4    l_local := 'abcxyz';
  5    dbms_output.put_line(l_local || chr(10));
  6  end;
  7  /

PL/SQL procedure successfully completed.

SQL> select sql_text from v$sql where sql_text like '%abcxyz%';
SQL_TEXT      
                                                                                                                                                                                                                                  
DECLARE 
SqlDevBind1Z_1 VARCHAR2(32767):=:SqlDevBind1ZInit1;  
BEGIN 
   declare   
       l_local varchar2(20); 
   begin   
       l_local := 'abcxyz';   
       dbms_output.put_line(l_local || chr(TO_NUMBER( SqlDevBind1Z_1))); 
   end;  
 :AUXSQLDBIND1:=SqlDevBind1Z_1;  
END;  



New Performance


So I spent the better part of 2 week in the NetBeans profiler and the outcome is well worth the time.  ALL these numbers are on my laptop so milage will vary.  APEX is probably one of the largest / complicated set of sql / plsql scripts to install into a database so I used that as my baseline. The SQLcl version I started from took 19m27.352s to install APEX.  For comparison, I ran the same install with SQL*PLUS that took almost 10 full minutes less at 9m59.789s.  SOOOO clearly there was an issue here.

The key thing is knowing WHAT your application should be doing and how it should be doing it.  There were a number of things that SQLcl was being overly aggressive  about such as securing literals  which was introduced here http://krisrice.blogspot.com/2015/09/sqlcl-more-secure-now-with-rest.html Then there were calls that were repetitive and could simple be removed.  Then the next boost was from being more lax on dbms_output.  SQLcl was checking for things like DDL that clearly can't have output so no need to check.

The end result is that turned secure literals off and it now takes on my machine 7m17.635s