Can I control how often the scheduler wakes up?

Question

We're experimenting with parallelizing long running jobs by breaking them up into smaller mutually exclusive jobs. The majority of the small jobs run in a subsecond timeframe. When I send these jobs to the job queue via dbms_job, I get respectable throughput -- along the lines of hundreds per second. When I converted this to run with dbms_scheduler, the throughput is horrible. It will run 12 jobs in subsecond timeframe but then not run the next set for 5 seconds (run 12, 5 seconds later run 12 more, etc.) even though there are many jobs queued up to run (hundreds). I seem to remember a way to govern how long the job queue processor "sleeps" in the dbms_job scenario although I can't seem to find it in the 10g documentation (going to check the 9i documentation next). Is there a way to configure the job scheduler similarly as to how long it should sleep after it runs a set of jobs?

Answer 1

What version are you running?
There are 2 reasons why this is happening:
1)
To is designed to run jobs in max throughput, i.e it will throttle jobs to let the ones
that are running finish quicker. So it is a little conservative in scheduling jobs in parallel.
2) Scheduler works in conjunction with resource manager. Resource manager
throttles resource groups. If you have don't have resource manager running you will
effectively have a single consumer group. This also limits the number of jobs you can run in parallel.
To run many jobs in parallel you need to make a resource plan with many resource groups.
Example:
The following is an example how you can get 80 slaves to run (assuming cpu_count = 2)
alter system set resource_manager_plan = '';
grant create job to scott;
Rem Drop objects if they exist
connect scott/tiger
BEGIN
   for i in 0..19 loop
   for j in 0..3 loop       /* 2 + cpu_count jobs per class */
   begin
   DBMS_SCHEDULER.DROP_JOB(
      JOB_NAME => 'MANY_ACTIVE_JOBS_'||i||'_'||j);
   exception when others then null;
   end;
   end loop;
   end loop;
END;
connect / as sysdba
BEGIN
   for i in 0..19 loop
   DBMS_SCHEDULER.DROP_JOB_CLASS(
      JOB_CLASS_NAME => 'MANY_ACTIVE_JOBS_CLASS_'||i);
   end loop;
END;
BEGIN
   DBMS_RESOURCE_MANAGER.CLEAR_PENDING_AREA();
   DBMS_RESOURCE_MANAGER.CREATE_PENDING_AREA();
   DBMS_RESOURCE_MANAGER.DELETE_PLAN_CASCADE ('MANY_JOBS');
   DBMS_RESOURCE_MANAGER.VALIDATE_PENDING_AREA();
   DBMS_RESOURCE_MANAGER.SUBMIT_PENDING_AREA();
END;
begin
   DBMS_RESOURCE_MANAGER.CLEAR_PENDING_AREA();
   DBMS_RESOURCE_MANAGER.CREATE_PENDING_AREA();
  for i in 0..19 loop
  begin
    DBMS_RESOURCE_MANAGER.DELETE_CONSUMER_GROUP(
     CONSUMER_GROUP=>'many_active_jobs_group_'||i);
   exception when others then null;
  end;
  end loop;
  DBMS_RESOURCE_MANAGER.SUBMIT_PENDING_AREA;
end;
commit;
Rem --- END DROP OBJECTS
Rem -- CREATE OBJECTS
begin
   DBMS_RESOURCE_MANAGER.CLEAR_PENDING_AREA();
   DBMS_RESOURCE_MANAGER.CREATE_PENDING_AREA();
  for i in 0..19 loop
  begin
    DBMS_RESOURCE_MANAGER.CREATE_CONSUMER_GROUP(
     CONSUMER_GROUP=>'many_active_jobs_group_'||i, comment=>'consumer group for job scheduling many active jobs');
   exception when others then null;
  end;
  end loop;
  DBMS_RESOURCE_MANAGER.SUBMIT_PENDING_AREA;
end;
commit;
BEGIN
   DBMS_RESOURCE_MANAGER.CLEAR_PENDING_AREA();
   DBMS_RESOURCE_MANAGER.CREATE_PENDING_AREA();
   DBMS_RESOURCE_MANAGER.CREATE_PLAN
      ('MANY_JOBS', 'PLAN TO SUPPORT MANY ACTIVE JOBS');
   for i in 0..19 loop
   DBMS_RESOURCE_MANAGER.CREATE_PLAN_DIRECTIVE(
      PLAN => 'MANY_JOBS',
      GROUP_OR_SUBPLAN => 'many_active_jobs_group_' || i,
      COMMENT=>'none',
      CPU_P1 => 5
   end loop;
   DBMS_RESOURCE_MANAGER.CREATE_PLAN_DIRECTIVE(
      PLAN => 'MANY_JOBS',
      GROUP_OR_SUBPLAN => 'OTHER_GROUPS',
      CPU_P1 => 0,
      CPU_P2 => 100,
      COMMENT=>'Required'
   DBMS_RESOURCE_MANAGER.VALIDATE_PENDING_AREA();
   DBMS_RESOURCE_MANAGER.SUBMIT_PENDING_AREA();
END;
BEGIN
   for i in 0..19 loop
   DBMS_SCHEDULER.CREATE_JOB_CLASS(
      JOB_CLASS_NAME => 'MANY_ACTIVE_JOBS_CLASS_'||i,
      LOGGING_LEVEL => DBMS_SCHEDULER.LOGGING_FULL,
      LOG_HISTORY => 5,
      RESOURCE_CONSUMER_GROUP => 'MANY_ACTIVE_JOBS_GROUP_'||i
  end loop;
END;
BEGIN
  for i in 0..19 loop
  DBMS_RESOURCE_MANAGER_PRIVS.GRANT_SWITCH_CONSUMER_GROUP(
      'SCOTT','MANY_ACTIVE_JOBS_GROUP_'||i, TRUE);
  end loop;
END;
alter system set resource_manager_plan = 'MANY_JOBS';
connect scott/tiger
create or replace procedure call_fib(n number) is
i number;
begin
    if n = 1 then null; return; end if;
    if n = 2 then null; return; end if;
    if n = 20 then select 0 into i from dual; end if;
    call_fib(n-1);
    call_fib(n-2);
end;
show errors procedure call_fib;
BEGIN
   for i in 0..19 loop
   for j in 0..3 loop
   DBMS_SCHEDULER.CREATE_JOB(JOB_NAME => 'MANY_ACTIVE_JOBS_' || i || '_'||j,
      JOB_TYPE=>'PLSQL_BLOCK', JOB_ACTION=>'call_fib(23);',
      JOB_CLASS=>'MANY_ACTIVE_JOBS_CLASS_'||i,
      REPEAT_INTERVAL=>'FREQ=SECONDLY',
      ENABLED=>true);
  end loop;
  end loop;
END;
/