COMMIT_WAIT and COMMIT_LOGGING

Recently I used the COMMIT_WAIT and COMMIT_LOGGING parameters for solving (or, better, working around) a problem I faced while optimizing a specific task for one of my customers. Since it was the first time I used them in a production system, I thought to write this post not only to shortly explain the purpose of the these two parameters, but also to show a case where it is sensible to use them.

The purpose of the two parameters is the following:

COMMIT_WAIT

• Simply put this parameter specifies whether a server process that issues a commit waits for the log writer while it writes the redo data in the redo log files.
• If it’s set to WAIT, the default value, the server process waits. And, in most situations, this is the best thing to do.
• If it’s set to NOWAIT, it doesn’t wait. This means that, in case of a crash just after a commit, the D of ACID might be violated! Hence, in general, it is not advised to use this value.
• If it’s set to FORCE_WAIT the behaviour is similar to WAIT. The only difference is that settings at a lower level are ignored. In other words, if it is set at the system level it overrides the setting at the session and transaction level. If it’s set at session level it overrides the setting at the transaction level.

COMMIT_LOGGING

• Simply put this parameter specifies whether the log writer writes redo data in batches.
• If it’s set to IMMEDIATE, the default value, it basically performs a write operation for each commit.
• If it’s set to BATCH, it writes redo data in batches. Since with this value less but larger write operations might be performed, in case of small transactions the log writer should be able to write the redo data in a more efficient way.

Note that in 10.2, the version that introduced these features, there is a single parameter (COMMIT_WRITE) to control them. For example, it is possible to set it to “NOWAIT, BATCH”.

To illustrates how these two parameters work I wrote two scripts: commit.sh and commit.sql. Their purpose is to show the number of times specific system calls are executed by the log writer and a server process that executes the following PL/SQL block:

DECLARE
  l_dummy INTEGER;
BEGIN
  FOR i IN 1..1000
  LOOP
    INSERT INTO t VALUES (i, rpad('*',100,'*'));
    COMMIT;
    SELECT count(*) INTO l_dummy FROM dual;
  END LOOP;
END;

The two system calls I was interested in were the following:

• semtimedop(2): simply put this one is used by the server process to wait for the log writer
• io_submit(2): simply put this one is used by the log writer to write redo data in the redo log files

Let’s have a look the output generated by the scripts for three particular cases:

• COMMIT_WAIT = WAIT and COMMIT_LOGGING = IMMEDIATE: Notice that the server process executes 1005 times semtimedop(2) and that the log writer executes 1016 times io_submit(2). In other words, for both of them the number of executions is approximately the number of commits performed by the PL/SQL block.

   oracle@helicon:~/commit/ [DBM11203] ./commit.sh chris ian wait immediate 2> /dev/null

   ***** Server Process *****

   % time     seconds  usecs/call     calls    errors syscall
   ------ ----------- ----------- --------- --------- ----------------
   100.00    0.069561          69      1005         1 semtimedop
   ------ ----------- ----------- --------- --------- ----------------
   100.00    0.069561                  1005         1 total

   ***** Log Writer *****

   % time     seconds  usecs/call     calls    errors syscall
   ------ ----------- ----------- --------- --------- ----------------
   100.00    0.013919          14      1016           io_submit
   ------ ----------- ----------- --------- --------- ----------------
   100.00    0.013919                  1016           total

• COMMIT_WAIT = NOWAIT and COMMIT_LOGGING = IMMEDIATE: Notice that, for the server process, the executions of semtimedop(2) dropped to 5. No noticeable difference is observable for the log writer.

oracle@helicon:~/commit/ [DBM11203] ./commit.sh chris ian nowait immediate 2> /dev/null

  ***** Server Process *****

  % time     seconds  usecs/call     calls    errors syscall
  ------ ----------- ----------- --------- --------- ----------------
  100.00    0.002195         439         5         1 semtimedop
  ------ ----------- ----------- --------- --------- ----------------
  100.00    0.002195                     5         1 total

  ***** Log Writer *****

  % time     seconds  usecs/call     calls    errors syscall
  ------ ----------- ----------- --------- --------- ----------------
  100.00    0.010073          10      1015           io_submit
  ------ ----------- ----------- --------- --------- ----------------
  100.00    0.010073                  1015           total

• COMMIT_WAIT = NOWAIT and COMMIT_LOGGING = BATCH: Notice that this time not only the number of executions to io_submit(2) dropped to 15, but, in total, much less time was spent for writing the redo data (10*1015 >> 36*15).

   oracle@helicon:~/commit/ [DBM11203] ./commit.sh chris ian nowait batch 2> /dev/null

  ***** Server Process *****

  % time     seconds  usecs/call     calls    errors syscall
  ------ ----------- ----------- --------- --------- ----------------
  100.00    0.002132         533         4         1 semtimedop
  ------ ----------- ----------- --------- --------- ----------------
  100.00    0.002132                     4         1 total

  ***** Log Writer *****

  % time     seconds  usecs/call     calls    errors syscall
  ------ ----------- ----------- --------- --------- ----------------
  100.00    0.000533          36        15           io_submit
  ------ ----------- ----------- --------- --------- ----------------
  100.00    0.000533                    15           total

Now that I explained what the purpose of these parameters is, let me describe a case where I successfully used them.

Few weeks ago one of my customers migrated its DMS to a new version. During the migration data had to be moved from one database to another. Unfortunately the migration code was written to process data slow by slow. Hence, to speed up the processing, parallelization was added to the picture.

The following chart (you can click on it to increase its size) shows the load generated by 20 parallel processes with default values for COMMIT_WAIT and COMMIT_LOGGING. With the default configutation the system was able to process about 1000 “objects” per second. As you can see there were a lot of waits related to the commit wait class.

Since this was a controlled processing using COMMIT_WAIT was not considered a problem. So, we tested the very same load with COMMIT_WAIT set to NOWAIT and COMMIT_LOGGING set to BATCH. The throughput increased to about 1200 “object” per second. In other words, not dramatically. But, more importantly, as the following chart shows almost all waits in the commit wait class disappeared. This was very important because without that serialization taking place we were able to increase the number of parallel processes.

The following chart shows the load generated by 50 parallel processes. Again, almost no wait related to commits. With 50 processes the system was able to process about 2300 “objects” per second.

In summary, COMMIT_WAIT and COMMIT_LOGGING are not commonly used parameters but, in some specific situations, using them it might be beneficial to avoid wait events related to commits.