StarPU Handbook
|
To get an idea of what is happening, a lot of performance feedback is available, detailed in this chapter. The various informations should be checked for.
You can also use the Temanejo task debugger (see Using The Temanejo Task Debugger) to visualize the task graph more easily.
StarPU can use the FxT library (see https://savannah.nongnu.org/projects/fkt/) to generate traces with a limited runtime overhead.
You can get a tarball from http://download.savannah.gnu.org/releases/fkt/?C=M
Compiling and installing the FxT library in the $FXTDIR
path is done following the standard procedure:
$ ./configure --prefix=$FXTDIR $ make $ make install
In order to have StarPU to generate traces, StarPU should be configured with the option --with-fxt :
$ ./configure --with-fxt=$FXTDIR
Or you can simply point the PKG_CONFIG_PATH
to $FXTDIR/lib/pkgconfig
and pass --with-fxt to configure
When FxT is enabled, a trace is generated when StarPU is terminated by calling starpu_shutdown(). The trace is a binary file whose name has the form prof_file_XXX_YYY
where XXX
is the user name, and YYY
is the MPI id of the process that used StarPU (or 0 when running a sequential program). One can change the name of the file by setting the environnement variable STARPU_FXT_SUFFIX, its contents will be used instead of prof_file_XXX
. This file is saved in the /tmp/
directory by default, or by the directory specified by the environment variable STARPU_FXT_PREFIX.
The additional configure
option --enable-fxt-lock can be used to generate trace events which describes the locks behaviour during the execution. It is however very heavy and should not be used unless debugging StarPU's internal locking.
The environment variable STARPU_FXT_TRACE can be set to 0 to disable the generation of the prof_file_XXX_YYY
file.
When the FxT trace file prof_file_something
has been generated, it is possible to generate different trace formats by calling:
$ starpu_fxt_tool -i /tmp/prof_file_something
Or alternatively, setting the environment variable STARPU_GENERATE_TRACE to 1
before application execution will make StarPU do it automatically at application shutdown.
One can also set the environment variable STARPU_GENERATE_TRACE_OPTIONS to specify options, see starpu_fxt_tool –help
, for example:
$ export STARPU_GENERATE_TRACE=1 $ export STARPU_GENERATE_TRACE_OPTIONS="-no-acquire"
When running a MPI application, STARPU_GENERATE_TRACE will not work as expected (each node will try to generate trace files, thus mixing outputs...), you have to collect the trace files from the MPI nodes, and specify them all on the command starpu_fxt_tool
, for instance:
$ starpu_fxt_tool -i /tmp/prof_file_something*
By default, the generated trace contains all informations. To reduce the trace size, various -no-foo
options can be passed to starpu_fxt_tool
, see starpu_fxt_tool –help
.
One of the generated files is a trace in the Paje format. The file, located in the current directory, is named paje.trace
. It can be viewed with ViTE (http://vite.gforge.inria.fr/) a trace visualizing open-source tool. To open the file paje.trace
with ViTE, use the following command:
$ vite paje.trace
Tasks can be assigned a name (instead of the default unknown
) by filling the optional starpu_codelet::name, or assigning them a performance model. The name can also be set with the field starpu_task::name or by using STARPU_NAME when calling starpu_task_insert().
Tasks are assigned default colors based on the worker which executed them (green for CPUs, yellow/orange/red for CUDAs, blue for OpenCLs, red for MICs, ...). To use a different color for every type of task, one can specify the option -c
to starpu_fxt_tool
or in STARPU_GENERATE_TRACE_OPTIONS. Tasks can also be given a specific color by setting the field starpu_codelet::color or the starpu_task::color. Colors are expressed with the following format 0xRRGGBB
(e.g 0xFF0000
for red). See basic_examples/task_insert_color
for examples on how to assign colors.
To get statistics on the time spend in runtime overhead, one can use the statistics plugin of ViTE. In Preferences, select Plugins. In "States Type", select "Worker State". Then click on "Reload" to update the histogram. The red "Idle" percentages are due to lack of parallelism, while the brown "Overhead" and "Scheduling" percentages are due to the overhead of the runtime and of the scheduler.
To identify tasks precisely, the application can also set the field starpu_task::tag_id or setting STARPU_TAG_ONLY when calling starpu_task_insert(). The value of the tag will then show up in the trace.
One can also introduce user-defined events in the diagram thanks to the starpu_fxt_trace_user_event_string() function.
One can also set the iteration number, by just calling starpu_iteration_push() at the beginning of submission loops and starpu_iteration_pop() at the end of submission loops. These iteration numbers will show up in traces for all tasks submitted from there.
Coordinates can also be given to data with the starpu_data_set_coordinates() or starpu_data_set_coordinates_array() function. In the trace, tasks will then be assigned the coordinates of the first data they write to.
Traces can also be inspected by hand by using the tool fxt_print
, for instance:
$ fxt_print -o -f /tmp/prof_file_something
Timings are in nanoseconds (while timings as seen in ViTE are in milliseconds).
Another generated trace file is a task graph described using the DOT language. The file, created in the current directory, is named dag.dot
file in the current directory. It is possible to get a graphical output of the graph by using the graphviz
library:
$ dot -Tpdf dag.dot -o output.pdf
Another generated trace file gives details on the executed tasks. The file, created in the current directory, is named tasks.rec
. This file is in the recutils format, i.e. Field: value
lines, and empty lines to separate each task. This can be used as a convenient input for various ad-hoc analysis tools. By default it only contains information about the actual execution. Performance models can be obtained by running starpu_tasks_rec_complete
on it:
$ starpu_tasks_rec_complete tasks.rec tasks2.rec
which will add EstimatedTime
lines which contain the performance model-estimated time (in µs) for each worker starting from 0. Since it needs the performance models, it needs to be run the same way as the application execution, or at least with STARPU_HOSTNAME
set to the hostname of the machine used for execution, to get the performance models of that machine.
Another possibility is to obtain the performance models as an auxiliary perfmodel.rec
file, by using the starpu_perfmodel_recdump
utility:
$ starpu_perfmodel_recdump tasks.rec -o perfmodel.rec
Another generated trace file is an activity trace. The file, created in the current directory, is named activity.data
. A profile of the application showing the activity of StarPU during the execution of the program can be generated:
$ starpu_workers_activity activity.data
This will create a file named 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 picture activity.eps
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.
When using modular schedulers (i.e. schedulers which use a modular architecture, and whose name start with "modular-"), the call to starpu_fxt_tool
will also produce a trace.html
file which can be viewed in a javascript-enabled web browser. It shows the flow of tasks between the components of the modular scheduler.
For computing statistics, it is useful to limit the trace to a given portion of the time of the whole execution. This can be achieved by calling
before calling starpu_init(), to prevent tracing from starting immediately. Then
and
can be used around the portion of code to be traced. This will show up as marks in the trace, and states of workers will only show up for that portion.
The performance model of codelets (see Performance Model Example) can be examined by using the tool starpu_perfmodel_display
:
$ 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>
Here, the codelets of the example lu
are available. We can examine the performance of the kernel 22
(in micro-seconds), which is history-based:
$ 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
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:
$ starpu_perfmodel_display -s non_linear_memset_regression_based 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 ...
The same can also be achieved by using StarPU's library API, see Performance Model and notably the function starpu_perfmodel_load_symbol(). The source code of the tool starpu_perfmodel_display
can be a useful example.
An XML output can also be printed by using the -x
option:
tools/starpu_perfmodel_display -x -s non_linear_memset_regression_based <?xml version="1.0" encoding="UTF-8"?> <!DOCTYPE StarPUPerfmodel SYSTEM "starpu-perfmodel.dtd"> <!-- symbol non_linear_memset_regression_based --> <!-- All times in us --> <perfmodel version="45"> <combination> <device type="CPU" id="0" ncores="1"/> <implementation id="0"> <!-- cpu0_impl0 (Comb0) --> <!-- time = a size ^b + c --> <nl_regression a="5.429195e-04" b="8.654899e-01" c="9.009313e-01"/> <entry footprint="a3d3725e" size="4096" flops="0.000000e+00" mean="4.763200e+00" deviation="7.650928e-01" nsample="100"/> <entry footprint="870a30aa" size="8192" flops="0.000000e+00" mean="1.827970e+00" deviation="2.037181e-01" nsample="100"/> <entry footprint="48e988e9" size="16384" flops="0.000000e+00" mean="2.652800e+00" deviation="1.876459e-01" nsample="100"/> <entry footprint="961e65d2" size="32768" flops="0.000000e+00" mean="4.255530e+00" deviation="3.518025e-01" nsample="100"/> </implementation> </combination> </perfmodel>
The tool starpu_perfmodel_plot
can be used to draw performance models. It writes a .gp
file in the current directory, to be run with the tool gnuplot
, which shows the corresponding curve.
When the field starpu_task::flops is set (or STARPU_FLOPS is passed to starpu_task_insert()), starpu_perfmodel_plot
can directly draw a GFlops/s curve, by simply adding the -f
option:
$ starpu_perfmodel_plot -f -s chol_model_11
This will however disable displaying the regression model, for which we can not compute GFlops/s.
When the FxT trace file prof_file_something
has been generated, it is possible to get a profiling of each codelet by calling:
$ starpu_fxt_tool -i /tmp/prof_file_something $ starpu_codelet_profile distrib.data codelet_name
This will create profiling data files, and a distrib.data.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 tool starpu_perfmodel_plot
, by passing it the fxt trace:
$ starpu_perfmodel_plot -s non_linear_memset_regression_based -i /tmp/prof_file_foo_0
It will produce a .gp
file which contains both the performance model curves, and the profiling measurements.
If you have the statistical tool R
installed, you can additionally use
$ starpu_codelet_histo_profile distrib.data
Which will create one .pdf
file per codelet and per input size, showing a histogram of the codelet execution time distribution.
It is possible to get statistics about tasks length and data size by using :
$ starpu_fxt_data_trace filename [codelet1 codelet2 ... codeletn]
Where filename is the FxT trace file and codeletX the names of the codelets you want to profile (if no names are specified, starpu_fxt_data_trace
will profile them all). This will create a file, data_trace.gp
which can be executed to get a .eps
image of these results. On the image, each point represents a task, and each color corresponds to a codelet.
More than just codelet performance, it is interesting to get statistics over all kinds of StarPU states (allocations, data transfers, etc.). This is particularly useful to check what may have gone wrong in the accurracy of the SimGrid simulation.
This requires the R
statistical tool, with the plyr
, ggplot2
and data.table
packages. If your system distribution does not have packages for these, one can fetch them from CRAN
:
$ R > install.packages("plyr") > install.packages("ggplot2") > install.packages("data.table") > install.packages("knitr")
The pj_dump
tool from pajeng
is also needed (see https://github.com/schnorr/pajeng)
One can then get textual or .csv
statistics over the trace states:
$ starpu_paje_state_stats -v native.trace simgrid.trace "Value" "Events_native.csv" "Duration_native.csv" "Events_simgrid.csv" "Duration_simgrid.csv" "Callback" 220 0.075978 220 0 "chol_model_11" 10 565.176 10 572.8695 "chol_model_21" 45 9184.828 45 9170.719 "chol_model_22" 165 64712.07 165 64299.203 $ starpu_paje_state_stats native.trace simgrid.trace
An other way to get statistics of StarPU states (without installing R
and pj_dump
) is to use the starpu_trace_state_stats.py
script which parses the generated trace.rec
file instead of the paje.trace
file. The output is similar to the previous script but it doesn't need any dependencies.
The different prefixes used in trace.rec
are:
E: Event type N: Event name C: Event category W: Worker ID T: Thread ID S: Start time
Here's an example on how to use it:
$ starpu_trace_state_stats.py trace.rec | column -t -s "," "Name" "Count" "Type" "Duration" "Callback" 220 Runtime 0.075978 "chol_model_11" 10 Task 565.176 "chol_model_21" 45 Task 9184.828 "chol_model_22" 165 Task 64712.07
starpu_trace_state_stats.py
can also be used to compute the different efficiencies. Refer to the usage description to show some examples.
And one can plot histograms of execution times, of several states for instance:
$ starpu_paje_draw_histogram -n chol_model_11,chol_model_21,chol_model_22 native.trace simgrid.trace
and see the resulting pdf file:
A quick statistical report can be generated by using:
$ starpu_paje_summary native.trace simgrid.trace
it includes gantt charts, execution summaries, as well as state duration charts and time distribution histograms.
Other external Paje analysis tools can be used on these traces, one just needs to sort the traces by timestamp order (which not guaranteed to make recording more efficient):
$ starpu_paje_sort paje.trace
StarPU can 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.
Theoretical Lower Bound On Execution Time Example provides an example on how to use this.
For kernels with history-based performance models (and provided that they are completely calibrated), StarPU can very easily provide a theoretical lower bound for the execution time of a whole set of tasks. See for instance examples/lu/lu_example.c
: before submitting tasks, call the function starpu_bound_start(), and after complete execution, call starpu_bound_stop(). starpu_bound_print_lp() or starpu_bound_print_mps() can then be used to output a Linear Programming problem corresponding to the schedule of your tasks. Run it through lp_solve
or any other linear programming solver, and that will give you a lower bound for the total execution time of your tasks. If StarPU was compiled with the library glpk
installed, starpu_bound_compute() can be used to solve it immediately and get the optimized minimum, in ms. Its parameter integer
allows to decide whether integer resolution should be computed and returned
The deps
parameter tells StarPU whether to take tasks, implicit data, and tag dependencies into account. Tags released in a callback or similar are not taken into account, only tags associated with a task are. It must be understood that the linear programming problem size is quadratic with the number of tasks and thus the time to solve it will be very long, it could be minutes for just a few dozen tasks. You should probably use lp_solve -timeout 1 test.pl -wmps test.mps
to convert the problem to MPS format and then use a better solver, glpsol
might be better than lp_solve
for instance (the –pcost
option may be useful), but sometimes doesn't manage to converge. cbc
might look slower, but it is parallel. For lp_solve
, be sure to try at least all the -B
options. For instance, we often just use lp_solve -cc -B1 -Bb -Bg -Bp -Bf -Br -BG -Bd -Bs -BB -Bo -Bc -Bi
, and the -gr
option can also be quite useful. The resulting schedule can be observed by using the tool starpu_lp2paje
, which converts it into the Paje format.
Data transfer time can only be taken into account when deps
is set. Only data transfers inferred from implicit data dependencies between tasks are taken into account. Other data transfers are assumed to be completely overlapped.
Setting deps
to 0 will only take into account the actual computations on processing units. It however still properly takes into account the varying performances of kernels and processing units, which is quite more accurate than just comparing StarPU performances with the fastest of the kernels being used.
The prio
parameter tells StarPU whether to simulate taking into account the priorities as the StarPU scheduler would, i.e. schedule prioritized tasks before less prioritized tasks, to check to which extend this results to a less optimal solution. This increases even more computation time.
Creating views with StarVZ (see: https://github.com/schnorr/starvz) is made up of two steps. The initial stage consists of a pre-processing of the traces generated by the application, while the second one consists of the analysis itself and is carried out with R packages' aid. StarVZ is available at CRAN (https://cran.r-project.org/package=starvz) and depends on pj_dump (from pajeng) and rec2csv (from recutils).
To download and install StarVZ, it is necessary to have R, pajeng, and recutils:
# For pj_dump and rec2csv apt install -y pajeng recutils # For R apt install -y r-base libxml2-dev libssl-dev libcurl4-openssl-dev libgit2-dev libboost-dev
To install the StarVZ, the following command can be used:
echo "install.packages('starvz', repos = 'https://cloud.r-project.org')" | R --vanilla
To generate traces from an application, it is necessary to set STARPU_GENERATE_TRACE and build StarPU with FxT. Then, StarVZ can be used on a folder with StarPU FxT traces to produce a default view:
export PATH=$(Rscript -e 'cat(system.file("tools/", package = "starvz"), sep="\n")'):$PATH starvz /foo/path-to-fxt-files
An example of default view:
One can also use existing trace files (paje.trace, tasks.rec, data.rec, papi.rec and dag.dot) skipping the StarVZ internal call to starpu_fxt_tool with:
starvz --use-paje-trace /foo/path-to-trace-files
Alternatively, each StarVZ step can be executed separately. Step 1 can be used on a folder with:
starvz -1 /foo/path-to-fxt-files
Then the second step can be executed directly in R. StarVZ enables a set of different plots that can be configured on a .yaml file. A default file is provided (default.yaml
); also, the options can be changed directly in R.
library(starvz) library(dplyr) dtrace <- starvz_read("./", selective = FALSE) # show idleness ratio dtrace$config$st$idleness = TRUE # show ABE bound dtrace$config$st$abe$active = TRUE # find the last task with dplyr dtrace$config$st$tasks$list = dtrace$Application %>% filter(End == max(End)) %>% .$JobId # show last task dependencies dtrace$config$st$tasks$active = TRUE dtrace$config$st$tasks$levels = 50 plot <- starvz_plot(dtrace)
An example of visualization follows:
It is possible to enable memory statistics. To do so, you need to pass the option --enable-memory-stats when running configure
. It is then possible to call the function starpu_data_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 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:
$ 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
$ 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
Different data statistics can be displayed at the end of the execution of the application. To enable them, you need to define the environment variable STARPU_ENABLE_STATS. When calling 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 STARPU_STATS to 0
.
$ ./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 %) ...
$ STARPU_STATS=0 ./examples/cholesky/cholesky_tag Computation took (in ms) 518.16 Synthetic GFlop/s : 44.21