[comment {-*- tcl -*- doctools manpage}]
[manpage_begin ttrace n 2.8]
[moddesc {Tcl Threading}]
[titledesc {Trace-based interpreter initialization}]
[require Tcl 8.4]
[require Thread [opt 2.8]]

[description]
This package creates a framework for on-demand replication of the
interpreter state across threads in an multithreading application.
It relies on the mechanics of Tcl command tracing and the Tcl
[cmd unknown] command and mechanism.
[para]
The package requires Tcl threading extension but can be alternatively
used stand-alone within the AOLserver, a scalable webserver from
America Online.
[para]
In a nutshell, a short sample illustrating the usage of the ttrace
with the Tcl threading extension:

[example {

    % package require Ttrace
    2.8.2

    % set t1 [thread::create {package require Ttrace; thread::wait}]
    tid0x1802800

    % ttrace::eval {proc test args {return test-[thread::id]}}
    % thread::send $t1 test
    test-tid0x1802800

    % set t2 [thread::create {package require Ttrace; thread::wait}]
    tid0x1804000

    % thread::send $t2 test
    test-tid0x1804000

}]
[para]
As seen from above, the [cmd ttrace::eval] and [cmd ttrace::update]
commands are used to create a thread-wide definition of a simple
Tcl procedure and replicate that definition to all, already existing
or later created, threads.

[section {USER COMMANDS}]
This section describes user-level commands. Those commands can be
used by script writers to control the execution of the tracing
framework.

[list_begin definitions]

[call [cmd ttrace::eval] [arg arg] [opt {arg ...}]]

This command concatenates given arguments and evaluates the resulting
Tcl command with trace framework enabled. If the command execution
was ok, it takes necessary steps to automatically propagate the
trace epoch change to all threads in the application.
For AOLserver, only newly created threads actually receive the
epoch change. For the Tcl threading extension, all threads created by
the extension are automatically updated. If the command execution
resulted in Tcl error, no state propagation takes place.
[para]
This is the most important user-level command of the package as
it wraps most of the commands described below. This greatly
simplifies things, because user need to learn just this (one)
command in order to effectively use the package. Other commands,
as described below, are included mostly for the sake of completeness.

[call [cmd ttrace::enable]]

Activates all registered callbacks in the framework
and starts a new trace epoch. The trace epoch encapsulates all
changes done to the interpreter during the time traces are activated.

[call [cmd ttrace::disable]]

Deactivates all registered callbacks in the framework
and closes the current trace epoch.

[call [cmd ttrace::cleanup]]

Used to clean-up all on-demand loaded resources in the interpreter.
It effectively brings Tcl interpreter to its pristine state.

[call [cmd ttrace::update] [opt epoch]]

Used to refresh the state of the interpreter to match the optional
trace [opt epoch]. If the optional [opt epoch] is not given, it takes
the most recent trace epoch.

[call [cmd ttrace::getscript]]

Returns a synthesized Tcl script which may be sourced in any interpreter.
This script sets the stage for the Tcl [cmd unknown] command so it can
load traced resources from the in-memory database. Normally, this command
is automatically invoked by other higher-level commands like
[cmd ttrace::eval] and [cmd ttrace::update].

[list_end]

[section {CALLBACK COMMANDS}]
A word upfront: the package already includes callbacks for tracing
following Tcl commands: [cmd proc], [cmd namespace], [cmd variable],
[cmd load], and [cmd rename]. Additionally, a set of callbacks for
tracing resources (object, classes) for the XOTcl v1.3.8+, an
OO-extension to Tcl, is also provided.
This gives a solid base for solving most of the real-life needs and
serves as an example for people wanting to customize the package
to cover their specific needs.
[para]
Below, you can find commands for registering callbacks in the
framework and for writing callback scripts. These callbacks are
invoked by the framework in order to gather interpreter state
changes, build in-memory database, perform custom-cleanups and
various other tasks.


[list_begin definitions]

[call [cmd ttrace::atenable] [arg cmd] [arg arglist] [arg body]]

Registers Tcl callback to be activated at [cmd ttrace::enable].
Registered callbacks are activated on FIFO basis. The callback
definition includes the name of the callback, [arg cmd], a list
of callback arguments, [arg arglist] and the [arg body] of the
callback. Effectively, this actually resembles the call interface
of the standard Tcl [cmd proc] command.


[call [cmd ttrace::atdisable] [arg cmd] [arg arglist] [arg body]]

Registers Tcl callback to be activated at [cmd ttrace::disable].
Registered callbacks are activated on FIFO basis. The callback
definition includes the name of the callback, [arg cmd], a list
of callback arguments, [arg arglist] and the [arg body] of the
callback. Effectively, this actually resembles the call interface
of the standard Tcl [cmd proc] command.


[call [cmd ttrace::addtrace] [arg cmd] [arg arglist] [arg body]]

Registers Tcl callback to be activated for tracing the Tcl
[cmd cmd] command. The callback definition includes the name of
the Tcl command to trace, [arg cmd], a list of callback arguments,
[arg arglist] and the [arg body] of the callback. Effectively,
this actually resembles the call interface of the standard Tcl
[cmd proc] command.


[call [cmd ttrace::addscript] [arg name] [arg body]]

Registers Tcl callback to be activated for building a Tcl
script to be passed to other interpreters. This script is
used to set the stage for the Tcl [cmd unknown] command.
Registered callbacks are activated on FIFO basis.
The callback definition includes the name of the callback,
[arg name] and the [arg body] of the callback.

[call [cmd ttrace::addresolver] [arg cmd] [arg arglist] [arg body]]

Registers Tcl callback to be activated by the overloaded Tcl
[cmd unknown] command.
Registered callbacks are activated on FIFO basis.
This callback is used to resolve the resource and load the
resource in the current interpreter.

[call [cmd ttrace::addcleanup] [arg body]]

Registers Tcl callback to be activated by the [cmd trace::cleanup].
Registered callbacks are activated on FIFO basis.

[call [cmd ttrace::addentry] [arg cmd] [arg var] [arg val]]

Adds one entry to the named in-memory database.

[call [cmd ttrace::getentry] [arg cmd] [arg var]]

Returns the value of the entry from the named in-memory database.

[call [cmd ttrace::getentries] [arg cmd] [opt pattern]]

Returns names of all entries from the named in-memory database.

[call [cmd ttrace::delentry] [arg cmd]]

Deletes an entry from the named in-memory database.

[call [cmd ttrace::preload] [arg cmd]]

Registers the Tcl command to be loaded in the interpreter.
Commands registered this way will always be the part of
the interpreter and not be on-demand loaded by the Tcl
[cmd unknown] command.

[list_end]

[section DISCUSSION]
Common introspective state-replication approaches use a custom Tcl
script to introspect the running interpreter and synthesize another
Tcl script to replicate this state in some other interpreter.
This package, on the contrary, uses Tcl command traces. Command
traces are registered on selected Tcl commands, like [cmd proc],
[cmd namespace], [cmd load] and other standard (and/or user-defined)
Tcl commands. When activated, those traces build an in-memory
database of created resources. This database is used as a resource
repository for the (overloaded) Tcl [cmd unknown] command which
creates the requested resource in the interpreter on demand.
This way, users can update just one interpreter (master) in one
thread and replicate that interpreter state (or part of it) to other
threads/interpreters in the process.
[para]
Immediate benefit of such approach is the much smaller memory footprint
of the application and much faster thread creation. By not actually
loading all necessary procedures (and other resources) in every thread
at the thread initialization time, but by deferring this to the time the
resource is actually referenced, significant improvements in both
memory consumption and thread initialization time can be achieved. Some
tests have shown that memory footprint of an multithreading Tcl application
went down more than three times and thread startup time was reduced for
about 50 times. Note that your mileage may vary.

Other benefits include much finer control about what (and when) gets
replicated from the master to other Tcl thread/interpreters.

[see_also tsv tpool thread]

[keywords {command tracing} introspection]

[manpage_end]