starpu-max/doc/chapters/perf-optimization.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  Centre National de la Recherche Scientifique
@c Copyright (C) 2011 Institut National de Recherche en Informatique et Automatique
@c See the file starpu.texi for copying conditions.

TODO: improve!

@menu
* Data management::
* Task granularity::
* Task submission::
* Task priorities::
* Task scheduling policy::
* Performance model calibration::
* Task distribution vs Data transfer::
* Data prefetch::
* Power-based scheduling::
* Profiling::
* CUDA-specific optimizations::
@end menu

Simply encapsulating application kernels into tasks already permits to
seamlessly support CPU and GPUs at the same time. To achieve good performance, a
few additional changes are needed.

@node Data management
@section Data management

When the application allocates data, whenever possible it should use the
@code{starpu_malloc} function, which will ask CUDA or
OpenCL to make the allocation itself and pin the corresponding allocated
memory. This is needed to permit asynchronous data transfer, i.e. permit data
transfer to overlap with computations. Otherwise, the trace will show that the
@code{DriverCopyAsync} state takes a lot of time, this is because CUDA or OpenCL
then reverts to synchronous transfers.

By default, StarPU leaves replicates of data wherever they were used, in case they
will be re-used by other tasks, thus saving the data transfer time. When some
task modifies some data, all the other replicates are invalidated, and only the
processing unit which ran that task will have a valid replicate of the data. If the application knows
that this data will not be re-used by further tasks, it should advise StarPU to
immediately replicate it to a desired list of memory nodes (given through a
bitmask). This can be understood like the write-through mode of CPU caches.

@cartouche
@smallexample
starpu_data_set_wt_mask(img_handle, 1<<0);
@end smallexample
@end cartouche

will for instance request to always automatically transfer a replicate into the
main memory (node 0), as bit 0 of the write-through bitmask is being set.

@cartouche
@smallexample
starpu_data_set_wt_mask(img_handle, ~0U);
@end smallexample
@end cartouche

will request to always automatically broadcast the updated data to all memory
nodes.

@node Task granularity
@section Task granularity

Like any other runtime, StarPU has some overhead to manage tasks. Since
it does smart scheduling and data management, that overhead is not always
neglectable. The order of magnitude of the overhead is typically a couple of
microseconds. The amount of work that a task should do should thus be somewhat
bigger, to make sure that the overhead becomes neglectible. The offline
performance feedback can provide a measure of task length, which should thus be
checked if bad performance are observed.

@node Task submission
@section Task submission

To let StarPU make online optimizations, tasks should be submitted
asynchronously as much as possible. Ideally, all the tasks should be
submitted, and mere calls to @code{starpu_task_wait_for_all} or
@code{starpu_data_unregister} be done to wait for
termination. StarPU will then be able to rework the whole schedule, overlap
computation with communication, manage accelerator local memory usage, etc.

@node Task priorities
@section Task priorities

By default, StarPU will consider the tasks in the order they are submitted by
the application. If the application programmer knows that some tasks should
be performed in priority (for instance because their output is needed by many
other tasks and may thus be a bottleneck if not executed early enough), the
@code{priority} field of the task structure should be set to transmit the
priority information to StarPU.

@node Task scheduling policy
@section Task scheduling policy

By default, StarPU uses the @code{eager} simple greedy scheduler. This is
because it provides correct load balance even if the application codelets do not
have performance models. If your application codelets have performance models
(@pxref{Performance model example} for examples showing how to do it),
you should change the scheduler thanks to the @code{STARPU_SCHED} environment
variable. For instance @code{export STARPU_SCHED=dmda} . Use @code{help} to get
the list of available schedulers.

The @b{eager} scheduler uses a central task queue, from which workers draw tasks
to work on. This however does not permit to prefetch data since the scheduling
decision is taken late. If a task has a non-0 priority, it is put at the front of the queue.

The @b{prio} scheduler also uses a central task queue, but sorts tasks by
priority (between -5 and 5).

The @b{random} scheduler distributes tasks randomly according to assumed worker
overall performance.

The @b{ws} (work stealing) scheduler schedules tasks on the local worker by
default. When a worker becomes idle, it steals a task from the most loaded
worker.

The @b{dm} (deque model) scheduler uses task execution performance models into account to
perform an HEFT-similar scheduling strategy: it schedules tasks where their
termination time will be minimal.

The @b{dmda} (deque model data aware) scheduler is similar to dm, it also takes
into account data transfer time.

The @b{dmdar} (deque model data aware ready) scheduler is similar to dmda,
it also sorts tasks on per-worker queues by number of already-available data
buffers.

The @b{dmdas} (deque model data aware sorted) scheduler is similar to dmda, it
also supports arbitrary priority values.

The @b{heft} (HEFT) scheduler is similar to dmda, it also supports task bundles.

The @b{pheft} (parallel HEFT) scheduler is similar to heft, it also supports
parallel tasks (still experimental).

The @b{pgreedy} (parallel greedy) scheduler is similar to greedy, it also
supports parallel tasks (still experimental).

@node Performance model calibration
@section Performance model calibration

Most schedulers are based on an estimation of codelet duration on each kind
of processing unit. For this to be possible, the application programmer needs
to configure a performance model for the codelets of the application (see
@ref{Performance model example} for instance). History-based performance models
use on-line calibration.  StarPU will automatically calibrate codelets
which have never been calibrated yet, and save the result in
@code{~/.starpu/sampling/codelets}.
The models are indexed by machine name. To share the models between machines (e.g. for a homogeneous cluster), use @code{export STARPU_HOSTNAME=some_global_name}. To force continuing calibration, use
@code{export STARPU_CALIBRATE=1} . This may be necessary if your application
has not-so-stable performance. StarPU will force calibration (and thus ignore
the current result) until 10 (_STARPU_CALIBRATION_MINIMUM) measurements have been
made on each architecture, to avoid badly scheduling tasks just because the
first measurements were not so good. Details on the current performance model status
can be obtained from the @code{starpu_perfmodel_display} command: the @code{-l}
option lists the available performance models, and the @code{-s} option permits
to choose the performance model to be displayed. The result looks like:

@example
$ starpu_perfmodel_display -s starpu_dlu_lu_model_22
performance model for cpu
# hash    size     mean          dev           n
880805ba  98304    2.731309e+02  6.010210e+01  1240
b50b6605  393216   1.469926e+03  1.088828e+02  1240
5c6c3401  1572864  1.125983e+04  3.265296e+03  1240
@end example

Which shows that for the LU 22 kernel with a 1.5MiB matrix, the average
execution time on CPUs was about 11ms, with a 3ms standard deviation, over
1240 samples. It is a good idea to check this before doing actual performance
measurements.

A graph can be drawn by using the @code{starpu_perfmodel_plot}:

@example
$ starpu_perfmodel_plot -s starpu_dlu_lu_model_22
98304 393216 1572864 
$ gnuplot starpu_starpu_dlu_lu_model_22.gp
$ gv starpu_starpu_dlu_lu_model_22.eps
@end example

If a kernel source code was modified (e.g. performance improvement), the
calibration information is stale and should be dropped, to re-calibrate from
start. This can be done by using @code{export STARPU_CALIBRATE=2}.

Note: due to CUDA limitations, to be able to measure kernel duration,
calibration mode needs to disable asynchronous data transfers. Calibration thus
disables data transfer / computation overlapping, and should thus not be used
for eventual benchmarks. Note 2: history-based performance models get calibrated
only if a performance-model-based scheduler is chosen.

@node Task distribution vs Data transfer
@section Task distribution vs Data transfer

Distributing tasks to balance the load induces data transfer penalty. StarPU
thus needs to find a balance between both. The target function that the
@code{dmda} scheduler of StarPU
tries to minimize is @code{alpha * T_execution + beta * T_data_transfer}, where
@code{T_execution} is the estimated execution time of the codelet (usually
accurate), and @code{T_data_transfer} is the estimated data transfer time. The
latter is estimated based on bus calibration before execution start,
i.e. with an idle machine, thus without contention. You can force bus re-calibration by running
@code{starpu_calibrate_bus}. The beta parameter defaults to 1, but it can be
worth trying to tweak it by using @code{export STARPU_SCHED_BETA=2} for instance,
since during real application execution, contention makes transfer times bigger.
This is of course imprecise, but in practice, a rough estimation already gives
the good results that a precise estimation would give.

@node Data prefetch
@section Data prefetch

The @code{heft}, @code{dmda} and @code{pheft} scheduling policies perform data prefetch (see @ref{STARPU_PREFETCH}):
as soon as a scheduling decision is taken for a task, requests are issued to
transfer its required data to the target processing unit, if needeed, so that
when the processing unit actually starts the task, its data will hopefully be
already available and it will not have to wait for the transfer to finish.

The application may want to perform some manual prefetching, for several reasons
such as excluding initial data transfers from performance measurements, or
setting up an initial statically-computed data distribution on the machine
before submitting tasks, which will thus guide StarPU toward an initial task
distribution (since StarPU will try to avoid further transfers).

This can be achieved by giving the @code{starpu_data_prefetch_on_node} function
the handle and the desired target memory node.

@node Power-based scheduling
@section Power-based scheduling

If the application can provide some power performance model (through
the @code{power_model} field of the codelet structure), StarPU will
take it into account when distributing tasks. The target function that
the @code{dmda} scheduler minimizes becomes @code{alpha * T_execution +
beta * T_data_transfer + gamma * Consumption} , where @code{Consumption}
is the estimated task consumption in Joules. To tune this parameter, use
@code{export STARPU_SCHED_GAMMA=3000} for instance, to express that each Joule
(i.e kW during 1000us) is worth 3000us execution time penalty. Setting
@code{alpha} and @code{beta} to zero permits to only take into account power consumption.

This is however not sufficient to correctly optimize power: the scheduler would
simply tend to run all computations on the most energy-conservative processing
unit. To account for the consumption of the whole machine (including idle
processing units), the idle power of the machine should be given by setting
@code{export STARPU_IDLE_POWER=200} for 200W, for instance. This value can often
be obtained from the machine power supplier.

The power actually consumed by the total execution can be displayed by setting
@code{export STARPU_PROFILING=1 STARPU_WORKER_STATS=1} .

@node Profiling
@section Profiling

A quick view of how many tasks each worker has executed can be obtained by setting 
@code{export STARPU_WORKER_STATS=1} This is a convenient way to check that
execution did happen on accelerators without penalizing performance with
the profiling overhead.

A quick view of how much data transfers have been issued can be obtained by setting 
@code{export STARPU_BUS_STATS=1} .

More detailed profiling information can be enabled by using @code{export STARPU_PROFILING=1} or by
calling @code{starpu_profiling_status_set} from the source code.
Statistics on the execution can then be obtained by using @code{export
STARPU_BUS_STATS=1} and @code{export STARPU_WORKER_STATS=1} .
 More details on performance feedback are provided by the next chapter.

@node CUDA-specific optimizations
@section CUDA-specific optimizations

Due to CUDA limitations, StarPU will have a hard time overlapping its own
communications and the codelet computations if the application does not use a
dedicated CUDA stream for its computations. StarPU provides one by the use of
@code{starpu_cuda_get_local_stream()} which should be used by all CUDA codelet
operations. For instance:

@cartouche
@smallexample
func <<<grid,block,0,starpu_cuda_get_local_stream()>>> (foo, bar);
cudaStreamSynchronize(starpu_cuda_get_local_stream());
@end smallexample
@end cartouche

StarPU already does appropriate calls for the CUBLAS library.

Unfortunately, some CUDA libraries do not have stream variants of
kernels. That will lower the potential for overlapping.