mirror of
https://github.com/MariaDB/server.git
synced 2025-01-16 03:52:35 +01:00
5449b03bc8
If no problems are discovered in the next week, this will make the replication code ready for 4.0.2 release. dbug/dbug.c: cleanup of my previous fix sql/mysqld.cc: fixed a REALLY NASTY BUG - slave threads were being launched before initialization of global thread keys. Thus if the slave thread was slow to start everything worked fine, but if it started quickly, we would get into trouble using the unitinialized keys sql/net_pkg.cc: make net_printf() work with 0 error code taking the third argument as format string in that case sql/slave.cc: misc fix-ups and debugging instrumentations sql/slave.h: added skip_log_purge member sql/sql_class.cc: debugging instrumentation to track down random memory corruption sql/sql_class.h: added debugging sentry to THD to track down memory corruption sql/sql_repl.cc: fixed bugs in CHANGE MASTER |
||
|---|---|---|
| .. | ||
| .cvsignore | ||
| dbug.c | ||
| dbug_analyze.c | ||
| dbug_long.h | ||
| doinstall.sh | ||
| example1.c | ||
| example2.c | ||
| example3.c | ||
| factorial.c | ||
| install.sh | ||
| main.c | ||
| Makefile.am | ||
| mklintlib.sh | ||
| monty.doc | ||
| qmake.cmd | ||
| readme.prof | ||
| sanity.c | ||
| user.r | ||
| vargs.h | ||
Hi,
I'm sending you the modifications I made to your Dbug routines to
allow profiling in a (relatively) machine independent fashion.
I use your Dbug routines fairly extensively. Unfortunately, it's
a royal pain to have to keep profiled versions of various libraries
around. The modifications allow profiling without the need for this.
How it works.
------------
Basically, I just added code in the dbug routines to write out a file
called dbugmon.out (by default). This is an ascii file containing lines
of the form:
<function-name> E <time-entered>
<function-name> X <time-exited>
A second program (analyze) reads this file, and produces a report on
standard output.
Profiling is enabled through the `g' flag. It can take a list of
procedure names for which profiling is enabled. By default, it
profiles all procedures.
The code in ``dbug.c'' opens the profile file for appending. This
is in order that one can run a program several times, and get the
sum total of all the times, etc.
The only system dependent part that I'm aware of is the routine
Clock() at the end of dbug.c. This returns the elapsed user time
in milliseconds. The version which I have is for 4.3 BSD. As I
don't have access to other systems, I'm not certain how this would
change.
An example of the report generated follows:
Profile of Execution
Execution times are in milliseconds
Calls Time
----- ----
Times Percentage Time Spent Percentage
Function Called of total in Function of total Importance
======== ====== ========== =========== ========== ==========
factorial 5 83.33 30 100.00 8333
main 1 16.67 0 0.00 0
======== ====== ========== =========== ==========
Totals 6 100.00 30 100.00
As you can see, it's quite self-evident. The ``Importance'' column is a
metric obtained by multiplying the percentage of the calls and the percentage
of the time. Functions with higher 'importance' benefit the most from
being sped up.
I'm really not certain how to add support for setjmp/longjmp, or for
child processes, so I've ignored that for the time being. In most of
the code that I write, it isn't necessary. If you have any good ideas,
feel free to add them.
This has been very useful to me. If you can use it as part of your
dbug distribution, please feel free to do so.
Regards,
Binayak Banerjee
{allegra | astrovax | bpa | burdvax}!sjuvax!bbanerje
bbanerje%sjuvax.sju.edu@relay.cs.net
July 9, 1987