starpu-max/doc/chapters/perf-feedback.texi

@c -*-texinfo-*-

@c This file is part of the StarPU Handbook.
@c Copyright (C) 2009--2011  Universit@'e de Bordeaux 1
@c Copyright (C) 2010, 2011, 2012, 2013  Centre National de la Recherche Scientifique
@c Copyright (C) 2011, 2012 Institut National de Recherche en Informatique et Automatique
@c See the file starpu.texi for copying conditions.

@menu
* Task debugger::               Using the Temanejo task debugger
* On-line::                     On-line performance feedback
* Off-line::                    Off-line performance feedback
* Codelet performance::         Performance of codelets
* Theoretical lower bound on execution time API::
* Memory feedback::
* Data statistics::
@end menu

@node Task debugger
@section Using the Temanejo task debugger

StarPU can connect to Temanejo (see
@url{http://www.hlrs.de/temanejo}), to permit
nice visual task debugging. To do so, build Temanejo's @code{libayudame.so},
install @code{Ayudame.h} to e.g. @code{/usr/local/include}, apply the
@code{tools/patch-ayudame} to it to fix C build, re-@code{./configure}, make
sure that it found it, rebuild StarPU.  Run the Temanejo GUI, give it the path
to your application, any options you want to pass it, the path to libayudame.so.

Make sure to specify at least the same number of CPUs in the dialog box as your
machine has, otherwise an error will happen during execution. Future versions
of Temanejo should be able to tell StarPU the number of CPUs to use.

Tag numbers have to be below @code{4000000000000000000ULL} to be usable for
Temanejo (so as to distinguish them from tasks).

@node On-line
@section On-line performance feedback

@menu
* Enabling on-line performance monitoring::
* Task feedback::               Per-task feedback
* Codelet feedback::            Per-codelet feedback
* Worker feedback::             Per-worker feedback
* Bus feedback::                Bus-related feedback
* StarPU-Top::                  StarPU-Top interface
@end menu

@node Enabling on-line performance monitoring
@subsection Enabling on-line performance monitoring

In order to enable online performance monitoring, the application can call
@code{starpu_profiling_status_set(STARPU_PROFILING_ENABLE)}. It is possible to
detect whether monitoring is already enabled or not by calling
@code{starpu_profiling_status_get()}. Enabling monitoring also reinitialize all
previously collected feedback. The @code{STARPU_PROFILING} environment variable
can also be set to 1 to achieve the same effect.

Likewise, performance monitoring is stopped by calling
@code{starpu_profiling_status_set(STARPU_PROFILING_DISABLE)}. Note that this
does not reset the performance counters so that the application may consult
them later on.

More details about the performance monitoring API are available in section
@ref{Profiling API}.

@node Task feedback
@subsection Per-task feedback

If profiling is enabled, a pointer to a @code{starpu_task_profiling_info}
structure is put in the @code{.profiling_info} field of the @code{starpu_task}
structure when a task terminates.
This structure is automatically destroyed when the task structure is destroyed,
either automatically or by calling @code{starpu_task_destroy}.

The @code{starpu_task_profiling_info} structure indicates the date when the
task was submitted (@code{submit_time}), started (@code{start_time}), and
terminated (@code{end_time}), relative to the initialization of
StarPU with @code{starpu_init}. It also specifies the identifier of the worker
that has executed the task (@code{workerid}).
These date are stored as @code{timespec} structures which the user may convert
into micro-seconds using the @code{starpu_timing_timespec_to_us} helper
function.

It it worth noting that the application may directly access this structure from
the callback executed at the end of the task. The @code{starpu_task} structure
associated to the callback currently being executed is indeed accessible with
the @code{starpu_task_get_current()} function.

@node Codelet feedback
@subsection Per-codelet feedback

The @code{per_worker_stats} field of the @code{struct starpu_codelet} structure is
an array of counters. The i-th entry of the array is incremented every time a
task implementing the codelet is executed on the i-th worker.
This array is not reinitialized when profiling is enabled or disabled.

@node Worker feedback
@subsection Per-worker feedback

The second argument returned by the @code{starpu_worker_get_profiling_info}
function is a @code{starpu_worker_profiling_info} structure that gives
statistics about the specified worker. This structure specifies when StarPU
started collecting profiling information for that worker (@code{start_time}),
the duration of the profiling measurement interval (@code{total_time}), the
time spent executing kernels (@code{executing_time}), the time spent sleeping
because there is no task to execute at all (@code{sleeping_time}), and the
number of tasks that were executed while profiling was enabled.
These values give an estimation of the proportion of time spent do real work,
and the time spent either sleeping because there are not enough executable
tasks or simply wasted in pure StarPU overhead.

Calling @code{starpu_worker_get_profiling_info} resets the profiling
information associated to a worker.

When an FxT trace is generated (see @ref{Generating traces}), it is also
possible to use the @code{starpu_workers_activity} script (described in @ref{starpu-workers-activity}) to
generate a graphic showing the evolution of these values during the time, for
the different workers.

@node Bus feedback
@subsection Bus-related feedback

TODO: ajouter STARPU_BUS_STATS

@c how to enable/disable performance monitoring

@c what kind of information do we get ?

The bus speed measured by StarPU can be displayed by using the
@code{starpu_machine_display} tool, for instance:

@example
StarPU has found:
        3 CUDA devices
                CUDA 0 (Tesla C2050 02:00.0)
                CUDA 1 (Tesla C2050 03:00.0)
                CUDA 2 (Tesla C2050 84:00.0)
from    to RAM          to CUDA 0       to CUDA 1       to CUDA 2
RAM     0.000000        5176.530428     5176.492994     5191.710722
CUDA 0  4523.732446     0.000000        2414.074751     2417.379201
CUDA 1  4523.718152     2414.078822     0.000000        2417.375119
CUDA 2  4534.229519     2417.069025     2417.060863     0.000000
@end example

@node StarPU-Top
@subsection StarPU-Top interface

StarPU-Top is an interface which remotely displays the on-line state of a StarPU
application and permits the user to change parameters on the fly.

Variables to be monitored can be registered by calling the
@code{starpu_top_add_data_boolean}, @code{starpu_top_add_data_integer},
@code{starpu_top_add_data_float} functions, e.g.:

@cartouche
@smallexample
starpu_top_data *data = starpu_top_add_data_integer("mynum", 0, 100, 1);
@end smallexample
@end cartouche

The application should then call @code{starpu_top_init_and_wait} to give its name
and wait for StarPU-Top to get a start request from the user. The name is used
by StarPU-Top to quickly reload a previously-saved layout of parameter display.

@cartouche
@smallexample
starpu_top_init_and_wait("the application");
@end smallexample
@end cartouche

The new values can then be provided thanks to
@code{starpu_top_update_data_boolean}, @code{starpu_top_update_data_integer},
@code{starpu_top_update_data_float}, e.g.:

@cartouche
@smallexample
starpu_top_update_data_integer(data, mynum);
@end smallexample
@end cartouche

Updateable parameters can be registered thanks to @code{starpu_top_register_parameter_boolean}, @code{starpu_top_register_parameter_integer}, @code{starpu_top_register_parameter_float}, e.g.:

@cartouche
@smallexample
float alpha;
starpu_top_register_parameter_float("alpha", &alpha, 0, 10, modif_hook);
@end smallexample
@end cartouche

@code{modif_hook} is a function which will be called when the parameter is being modified, it can for instance print the new value:

@cartouche
@smallexample
void modif_hook(struct starpu_top_param *d) @{
    fprintf(stderr,"%s has been modified: %f\n", d->name, alpha);
@}
@end smallexample
@end cartouche

Task schedulers should notify StarPU-Top when it has decided when a task will be
scheduled, so that it can show it in its Gantt chart, for instance:

@cartouche
@smallexample
starpu_top_task_prevision(task, workerid, begin, end);
@end smallexample
@end cartouche

Starting StarPU-Top@footnote{StarPU-Top is started via the binary
@code{starpu_top}.} and the application can be done two ways:

@itemize
@item The application is started by hand on some machine (and thus already
waiting for the start event). In the Preference dialog of StarPU-Top, the SSH
checkbox should be unchecked, and the hostname and port (default is 2011) on
which the application is already running should be specified. Clicking on the
connection button will thus connect to the already-running application.
@item StarPU-Top is started first, and clicking on the connection button will
start the application itself (possibly on a remote machine). The SSH checkbox
should be checked, and a command line provided, e.g.:

@example
$ ssh myserver STARPU_SCHED=dmda ./application
@end example

If port 2011 of the remote machine can not be accessed directly, an ssh port bridge should be added:

@example
$ ssh -L 2011:localhost:2011 myserver STARPU_SCHED=dmda ./application
@end example

and "localhost" should be used as IP Address to connect to.
@end itemize

@node Off-line
@section Off-line performance feedback

@menu
* Generating traces::           Generating traces with FxT
* Gantt diagram::               Creating a Gantt Diagram
* DAG::                         Creating a DAG with graphviz
* starpu-workers-activity::     Monitoring activity
@end menu

@node Generating traces
@subsection Generating traces with FxT

StarPU can use the FxT library (see
@url{https://savannah.nongnu.org/projects/fkt/}) to generate traces
with a limited runtime overhead.

You can either get a tarball:
@example
$ wget http://download.savannah.gnu.org/releases/fkt/fxt-0.2.11.tar.gz
@end example

or use the FxT library from CVS (autotools are required):
@example
$ cvs -d :pserver:anonymous@@cvs.sv.gnu.org:/sources/fkt co FxT
$ ./bootstrap
@end example

Compiling and installing the FxT library in the @code{$FXTDIR} path is
done following the standard procedure:
@example
$ ./configure --prefix=$FXTDIR
$ make
$ make install
@end example

In order to have StarPU to generate traces, StarPU should be configured with
the @code{--with-fxt} option:
@example
$ ./configure --with-fxt=$FXTDIR
@end example

Or you can simply point the @code{PKG_CONFIG_PATH} to
@code{$FXTDIR/lib/pkgconfig} and pass @code{--with-fxt} to @code{./configure}

When FxT is enabled, a trace is generated when StarPU is terminated by calling
@code{starpu_shutdown()}). The trace is a binary file whose name has the form
@code{prof_file_XXX_YYY} where @code{XXX} is the user name, and
@code{YYY} is the pid of the process that used StarPU. This file is saved in the
@code{/tmp/} directory by default, or by the directory specified by
the @code{STARPU_FXT_PREFIX} environment variable.

@node Gantt diagram
@subsection Creating a Gantt Diagram

When the FxT trace file @code{filename} has been generated, it is possible to
generate a trace in the Paje format by calling:
@example
$ starpu_fxt_tool -i filename
@end example

Or alternatively, setting the @code{STARPU_GENERATE_TRACE} environment variable
to @code{1} before application execution will make StarPU do it automatically at
application shutdown.

This will create a @code{paje.trace} file in the current directory that
can be inspected with the @url{http://vite.gforge.inria.fr/, ViTE trace
visualizing open-source tool}.  It is possible to open the
@code{paje.trace} file with ViTE by using the following command:
@example
$ vite paje.trace
@end example

To get names of tasks instead of "unknown", fill the optional @code{name} field
of the codelets, or use a performance model for them.

In the MPI execution case, collect the trace files from the MPI nodes, and
specify them all on the @code{starpu_fxt_tool} command, for instance:

@smallexample
$ starpu_fxt_tool -i filename1 -i filename2
@end smallexample

By default, all tasks are displayed using a green color. To display tasks with
varying colors, pass option @code{-c} to @code{starpu_fxt_tool}.

@node DAG
@subsection Creating a DAG with graphviz

When the FxT trace file @code{filename} has been generated, it is possible to
generate a task graph in the DOT format by calling:
@example
$ starpu_fxt_tool -i filename
@end example

This will create a @code{dag.dot} file in the current directory. This file is a
task graph described using the DOT language. It is possible to get a
graphical output of the graph by using the graphviz library:
@example
$ dot -Tpdf dag.dot -o output.pdf
@end example

@node starpu-workers-activity
@subsection Monitoring activity

When the FxT trace file @code{filename} has been generated, it is possible to
generate an activity trace by calling:
@example
$ starpu_fxt_tool -i filename
@end example

This will create an @code{activity.data} file in the current
directory. A profile of the application showing the activity of StarPU
during the execution of the program can be generated:
@example
$ starpu_workers_activity activity.data
@end example

This will create a file named @code{activity.eps} in the current directory.
This picture is composed of two parts.
The first part shows the activity of the different workers. The green sections
indicate which proportion of the time was spent executed kernels on the
processing unit. The red sections indicate the proportion of time spent in
StartPU: an important overhead may indicate that the granularity may be too
low, and that bigger tasks may be appropriate to use the processing unit more
efficiently. The black sections indicate that the processing unit was blocked
because there was no task to process: this may indicate a lack of parallelism
which may be alleviated by creating more tasks when it is possible.

The second part of the @code{activity.eps} picture is a graph showing the
evolution of the number of tasks available in the system during the execution.
Ready tasks are shown in black, and tasks that are submitted but not
schedulable yet are shown in grey.

@node Codelet performance
@section Performance of codelets

The performance model of codelets (described in @ref{Performance model example}) can be examined by using the
@code{starpu_perfmodel_display} tool:

@example
$ starpu_perfmodel_display -l
file: <malloc_pinned.hannibal>
file: <starpu_slu_lu_model_21.hannibal>
file: <starpu_slu_lu_model_11.hannibal>
file: <starpu_slu_lu_model_22.hannibal>
file: <starpu_slu_lu_model_12.hannibal>
@end example

Here, the codelets of the lu example are available. We can examine the
performance of the 22 kernel (in micro-seconds), which is history-based:

@example
$ starpu_perfmodel_display -s starpu_slu_lu_model_22
performance model for cpu
# hash      size       mean          dev           n
57618ab0    19660800   2.851069e+05  1.829369e+04  109
performance model for cuda_0
# hash      size       mean          dev           n
57618ab0    19660800   1.164144e+04  1.556094e+01  315
performance model for cuda_1
# hash      size       mean          dev           n
57618ab0    19660800   1.164271e+04  1.330628e+01  360
performance model for cuda_2
# hash      size       mean          dev           n
57618ab0    19660800   1.166730e+04  3.390395e+02  456
@end example

We can see that for the given size, over a sample of a few hundreds of
execution, the GPUs are about 20 times faster than the CPUs (numbers are in
us). The standard deviation is extremely low for the GPUs, and less than 10% for
CPUs.

This tool can also be used for regression-based performance models. It will then
display the regression formula, and in the case of non-linear regression, the
same performance log as for history-based performance models:

@example
$ starpu_perfmodel_display -s non_linear_memset_regression_based.type
performance model for cpu_impl_0
	Regression : #sample = 1400
	Linear: y = alpha size ^ beta
		alpha = 1.335973e-03
		beta = 8.024020e-01
	Non-Linear: y = a size ^b + c
		a = 5.429195e-04
		b = 8.654899e-01
		c = 9.009313e-01
# hash		size		mean		stddev		n
a3d3725e	4096           	4.763200e+00   	7.650928e-01   	100
870a30aa	8192           	1.827970e+00   	2.037181e-01   	100
48e988e9	16384          	2.652800e+00   	1.876459e-01   	100
961e65d2	32768          	4.255530e+00   	3.518025e-01   	100
...
@end example

The @code{starpu_perfmodel_plot} tool can be used to draw performance models.
It writes a @code{.gp} file in the current directory, to be run in the
@code{gnuplot} tool, which shows the corresponding curve.

The same can also be achieved by using StarPU's library API, see
@ref{Performance Model API} and notably the @code{starpu_perfmodel_load_symbol}
function. The source code of the @code{starpu_perfmodel_display} tool can be a
useful example.

When the FxT trace file @code{filename} has been generated, it is possible to
get a profiling of each codelet by calling:
@example
$ starpu_fxt_tool -i filename
$ starpu_codelet_profile distrib.data codelet_name
@end example

This will create profiling data files, and a @code{.gp} file in the current
directory, which draws the distribution of codelet time over the application
execution, according to data input size.

This is also available in the @code{starpu_perfmodel_plot} tool, by passing it
the fxt trace:

@example
$ starpu_perfmodel_display -s non_linear_memset_regression_based.type -i /tmp/prof_file_foo_0
@end example

It willd produce a @code{.gp} file which contains both the performance model
curves, and the profiling measurements.

If you have the R statistical tool installed, you can additionally use

@example
$ starpu_codelet_histo_profile distrib.data
@end example

Which will create one pdf file per codelet and per input size, showing a
histogram of the codelet execution time distribution.

@node Theoretical lower bound on execution time API
@section Theoretical lower bound on execution time

See @ref{Theoretical lower bound on execution time} for an example on how to use
this API. It permits to record a trace of what tasks are needed to complete the
application, and then, by using a linear system, provide a theoretical lower
bound of the execution time (i.e. with an ideal scheduling).

The computed bound is not really correct when not taking into account
dependencies, but for an application which have enough parallelism, it is very
near to the bound computed with dependencies enabled (which takes a huge lot
more time to compute), and thus provides a good-enough estimation of the ideal
execution time.

@deftypefun void starpu_bound_start (int @var{deps}, int @var{prio})
Start recording tasks (resets stats).  @var{deps} tells whether
dependencies should be recorded too (this is quite expensive)
@end deftypefun

@deftypefun void starpu_bound_stop (void)
Stop recording tasks
@end deftypefun

@deftypefun void starpu_bound_print_dot ({FILE *}@var{output})
Print the DAG that was recorded
@end deftypefun

@deftypefun void starpu_bound_compute ({double *}@var{res}, {double *}@var{integer_res}, int @var{integer})
Get theoretical upper bound (in ms) (needs glpk support detected by @code{configure} script). It returns 0 if some performance models are not calibrated.
@end deftypefun

@deftypefun void starpu_bound_print_lp ({FILE *}@var{output})
Emit the Linear Programming system on @var{output} for the recorded tasks, in
the lp format
@end deftypefun

@deftypefun void starpu_bound_print_mps ({FILE *}@var{output})
Emit the Linear Programming system on @var{output} for the recorded tasks, in
the mps format
@end deftypefun

@deftypefun void starpu_bound_print ({FILE *}@var{output}, int @var{integer})
Emit statistics of actual execution vs theoretical upper bound. @var{integer}
permits to choose between integer solving (which takes a long time but is
correct), and relaxed solving (which provides an approximate solution).
@end deftypefun

@node Memory feedback
@section Memory feedback

It is possible to enable memory statistics. To do so, you need to pass the option
@code{--enable-memory-stats} when running configure. It is then
possible to call the function @code{starpu_display_memory_stats()} to
display statistics about the current data handles registered within StarPU.

Moreover, statistics will be displayed at the end of the execution on
data handles which have not been cleared out. This can be disabled by
setting the environment variable @code{STARPU_MEMORY_STATS} to 0.

For example, if you do not unregister data at the end of the complex
example, you will get something similar to:

@example
$ STARPU_MEMORY_STATS=0 ./examples/interface/complex
Complex[0] = 45.00 + 12.00 i
Complex[0] = 78.00 + 78.00 i
Complex[0] = 45.00 + 12.00 i
Complex[0] = 45.00 + 12.00 i
@end example

@example
$ STARPU_MEMORY_STATS=1 ./examples/interface/complex
Complex[0] = 45.00 + 12.00 i
Complex[0] = 78.00 + 78.00 i
Complex[0] = 45.00 + 12.00 i
Complex[0] = 45.00 + 12.00 i

#---------------------
Memory stats:
#-------
Data on Node #3
#-----
Data : 0x553ff40
Size : 16

#--
Data access stats
/!\ Work Underway
Node #0
	Direct access : 4
	Loaded (Owner) : 0
	Loaded (Shared) : 0
	Invalidated (was Owner) : 0

Node #3
	Direct access : 0
	Loaded (Owner) : 0
	Loaded (Shared) : 1
	Invalidated (was Owner) : 0

#-----
Data : 0x5544710
Size : 16

#--
Data access stats
/!\ Work Underway
Node #0
	Direct access : 2
	Loaded (Owner) : 0
	Loaded (Shared) : 1
	Invalidated (was Owner) : 1

Node #3
	Direct access : 0
	Loaded (Owner) : 1
	Loaded (Shared) : 0
	Invalidated (was Owner) : 0
@end example

@node Data statistics
@section Data statistics

Different data statistics can be displayed at the end of the execution
of the application. To enable them, you need to pass the option
@code{--enable-stats} when calling @code{configure}. When calling
@code{starpu_shutdown()} various statistics will be displayed,
execution, MSI cache statistics, allocation cache statistics, and data
transfer statistics. The display can be disabled by setting the
environment variable @code{STARPU_STATS} to 0.

@example
$ ./examples/cholesky/cholesky_tag
Computation took (in ms)
518.16
Synthetic GFlops : 44.21
#---------------------
MSI cache stats :
TOTAL MSI stats	hit 1622 (66.23 %)	miss 827 (33.77 %)
...
@end example

@example
$ STARPU_STATS=0 ./examples/cholesky/cholesky_tag
Computation took (in ms)
518.16
Synthetic GFlops : 44.21
@end example

@c TODO: data transfer stats are similar to the ones displayed when
@c setting STARPU_BUS_STATS