Profiling Tcl

Profiling Tcl presents various techniques and resources for profiling the operational characteristics of a Tcl-based system. A good understanding of the timing, control flow, resource utilization, and code coverage of a process can help with debugging and optimization.


PracTcl Programming Tips by Stephen Uhler, Linux Journal, Issue 20, 1995-12
uses time. First rule of program optimization: Wait for a faster machine. Second rule: avoid optimizing code that doesn't matter.

Time It!

The simplest form of profiling is to just exploit Tcl's ability to time various commands, or to just print timestamps from within your application and perform the analysis manually. Procedures for this technique are outlined in How to Measure Performance.

Proc Profiling

You can redefine the proc command to insert timing code before and after each proc call. You can collect the data later from a global variable, and figure out which procedures are taking the most time. This is how the sage extension works, which you can find in the FAQ Extension Catalog [L1 ] (DEAD LINK) as:

 What: Sage
 Description: A Tcl/Tk runtime code analyzer profiling tool for Tcl/Tk
        Requires Tcl 8.0, and was tested on Linux i386 and SunOS/Solaris.
        Currently at version 1.1 .
 Updated: 11/2001
 Contact: mailto:[email protected]  (John Stump)

This adds overhead to each procedure call, so you will have to interpret the results. A short procedure, called millions of times, will suffer more from the overhead than a long procedure, called only a few times.

Command Profiling

If you want more detail, you could profile each and every Tcl command instead of just the procedures. This would add much more overhead than procedure profiling, but could give you lots more information.

If you combine both the procedure and command profiling, you should be able to get a sorted list of time consuming procedures, followed by the most time consuming commands within those procedures. You would be able to find out that the regexp command in your parse procedure was taking up the most time. Figuring out which regexp command (if there are several) is just a little more challenging. Put individual suspicious commands into their own procedures.

This combined procedure and command profiling is provided by the TclX extension, which is part of TclPro. This implementation is written in C instead of pure Tcl, so it is relatively fast. You can even start the profiling, run interactive commands, and save the results to a report file, like this:

> protclsh83
% package require Tclx
protclsh83.bin>profile -commands on
protclsh83.bin>source units.tcl
protclsh83.bin>source units.test
protclsh83.bin>profile off profarray
protclsh83.bin>profrep profarray cpu profdata.txt

The report would look something like this:

Procedure Call Stack                                Calls  Real Time    CPU Time
<global>                                                1      34125        2289
source                                                  3       2405        2258
units::convert                                         64       1826        1783
units::reduce                                         126       1815        1752
units::ReduceList                                      90       1627        1582
namespace                                               5        346         230
units::reduce                                          29        243         209
units::ReduceList                                      29        191         188

Line-by-Line Profiling

Instrumenting each line of code gives you the most detailed picture of program performance. You can easily tell which line within which procedure is creating the bottleneck. This technique is available for some compiled language profilers, but has not been implemented for Tcl (yet) (AFAIK -- RWT)

LV: Any techniques for profiling Tcl that doesn't involve modifying a Tcl application? Something that perhaps could be turned on in the interpreter itself?

RS 2008-07-16: Maybe Profiling with execution traces?


bll 2016-10-10 The quote in the reference section above "First rule of program optimization: Wait for a faster machine." is insanely stupid.

Or "Program runs slow? Throw more hardware at it."

This kind of attitude is one reason why websites are so slow nowadays. Faster hardware is not a replacement for poor program design.

See Also

Profiling Tcl with gprof
Profiling Tcl by Overloading Proc
(another) command profiler
a checkpoint-based profiler
A Look at the Tcl Test Suite with gcov