StarPU Handbook
|
TODO: improve!
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.
When the application allocates data, whenever possible it should use the function starpu_malloc(), 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 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.
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.
will request to always automatically broadcast the updated data to all memory nodes.
Setting the write-through mask to ~0U
can also be useful to make sure all memory nodes always have a copy of the data, so that it is never evicted when memory gets scarse.
Implicit data dependency computation can become expensive if a lot of tasks access the same piece of data. If no dependency is required on some piece of data (e.g. because it is only accessed in read-only mode, or because write accesses are actually commutative), use the function starpu_data_set_sequential_consistency_flag() to disable implicit dependencies on that data.
In the same vein, accumulation of results in the same data can become a bottleneck. The use of the mode STARPU_REDUX permits to optimize such accumulation (see Data Reduction).
Applications often need a data just for temporary results. In such a case, registration can be made without an initial value, for instance this produces a vector data:
StarPU will then allocate the actual buffer only when it is actually needed, e.g. directly on the GPU without allocating in main memory.
In the same vein, once the temporary results are not useful any more, the data should be thrown away. If the handle is not to be reused, it can be unregistered:
actual unregistration will be done after all tasks working on the handle terminate.
If the handle is to be reused, instead of unregistering it, it can simply be invalidated:
the buffers containing the current value will then be freed, and reallocated only when another task writes some value to the handle.
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, which is actually quite smaller than the CUDA overhead itself. 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. To get a grasp at the scalability possibility according to task size, one can run tests/microbenchs/tasks_size_overhead.sh
which draws curves of the speedup of independent tasks of very small sizes.
The choice of scheduler also has impact over the overhead: for instance, the scheduler dmda
takes time to make a decision, while eager
does not. tasks_size_overhead.sh
can again be used to get a grasp at how much impact that has on the target machine.
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 starpu_task_wait_for_all() or 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.
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 field starpu_task::priority should be set to transmit the priority information to StarPU.
The basics of the scheduling policy are that
push
operation) when they become ready to be executed, i.e. they are not waiting for some tags, data dependencies or task dependencies. pop
operation) one by one from the scheduler. This means scheduling policies usually contain at least one queue of tasks to store them between the time when they become available, and the time when a worker gets to grab them.
By default, StarPU uses the simple greedy scheduler eager
. 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 (Performance Model Example), you should change the scheduler thanks to the environment variable STARPU_SCHED. For instance export STARPU_SCHED=dmda
. Use help
to get the list of available schedulers.
The eager scheduler uses a central task queue, from which all workers draw tasks to work on concurrently. 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 prio scheduler also uses a central task queue, but sorts tasks by priority (between -5 and 5).
The random scheduler uses a queue per worker, and distributes tasks randomly according to assumed worker overall performance.
The ws (work stealing) scheduler uses a queue per worker, and schedules a task on the worker which released it by default. When a worker becomes idle, it steals a task from the most loaded worker.
The dm (deque model) scheduler uses task execution performance models into account to perform a HEFT-similar scheduling strategy: it schedules tasks where their termination time will be minimal. The difference with HEFT is that dm schedules tasks as soon as they become available, and thus in the order they become available, without taking priorities into account.
The dmda (deque model data aware) scheduler is similar to dm, but it also takes into account data transfer time.
The dmdar (deque model data aware ready) scheduler is similar to dmda, but it also sorts tasks on per-worker queues by number of already-available data buffers on the target device.
The dmdas (deque model data aware sorted) scheduler is similar to dmdar, except that it sorts tasks by priority order, which allows to become even closer to HEFT by respecting priorities after having made the scheduling decision (but it still schedules tasks in the order they become available).
The heft (heterogeneous earliest finish time) scheduler is a deprecated alias for dmda.
The pheft (parallel HEFT) scheduler is similar to dmda, it also supports parallel tasks (still experimental). Should not be used when several contexts using it are being executed simultaneously.
The peager (parallel eager) scheduler is similar to eager, it also supports parallel tasks (still experimental). Should not be used when several contexts using it are being executed simultaneously.
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 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 $STARPU_HOME/.starpu/sampling/codelets
. The models are indexed by machine name. To share the models between machines (e.g. for a homogeneous cluster), use export STARPU_HOSTNAME=some_global_name
. To force continuing calibration, use 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 command starpu_perfmodel_display
: the -l
option lists the available performance models, and the -s
option permits to choose the performance model to be displayed. The result looks like:
$ starpu_perfmodel_display -s starpu_slu_lu_model_11 performance model for cpu_impl_0 # hash size flops mean dev n 914f3bef 1048576 0.000000e+00 2.503577e+04 1.982465e+02 8 3e921964 65536 0.000000e+00 5.527003e+02 1.848114e+01 7 e5a07e31 4096 0.000000e+00 1.717457e+01 5.190038e+00 14 ...
Which shows that for the LU 11 kernel with a 1MiB matrix, the average execution time on CPUs was about 25ms, with a 0.2ms standard deviation, over 8 samples. It is a good idea to check this before doing actual performance measurements.
A graph can be drawn by using the tool starpu_perfmodel_plot
:
$ starpu_perfmodel_plot -s starpu_slu_lu_model_11 4096 16384 65536 262144 1048576 4194304 $ gnuplot starpu_starpu_slu_lu_model_11.gp $ gv starpu_starpu_slu_lu_model_11.eps
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 export STARPU_CALIBRATE=2
.
Note: history-based performance models get calibrated only if a performance-model-based scheduler is chosen.
The history-based performance models can also be explicitly filled by the application without execution, if e.g. the application already has a series of measurements. This can be done by using starpu_perfmodel_update_history(), for instance:
Measurement has to be provided in milliseconds for the completion time models, and in Joules for the energy consumption models.
Distributing tasks to balance the load induces data transfer penalty. StarPU thus needs to find a balance between both. The target function that the scheduler dmda
of StarPU tries to minimize is alpha * T_execution + beta * T_data_transfer
, where T_execution
is the estimated execution time of the codelet (usually accurate), and 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 the tool starpu_calibrate_bus
. The beta parameter defaults to 1
, but it can be worth trying to tweak it by using 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.
The scheduling policies heft
, dmda
and pheft
perform data prefetch (see 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 needed, 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 function starpu_data_prefetch_on_node() the handle and the desired target memory node.
If the application can provide some power performance model (through the field starpu_codelet::power_model), StarPU will take it into account when distributing tasks. The target function that the scheduler dmda
minimizes becomes alpha * T_execution + beta * T_data_transfer + gamma * Consumption
, where Consumption
is the estimated task consumption in Joules. To tune this parameter, use export STARPU_SCHED_GAMMA=3000
for instance, to express that each Joule (i.e kW during 1000us) is worth 3000us execution time penalty. Setting alpha
and 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 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 export STARPU_PROFILING=1 STARPU_WORKER_STATS=1
.
On-line task consumption measurement is currently only supported through the CL_PROFILING_POWER_CONSUMED
OpenCL extension, implemented in the MoviSim simulator. Applications can however provide explicit measurements by using the function starpu_perfmodel_update_history() (examplified in Performance Model Example with the power_model
performance model). Fine-grain measurement is often not feasible with the feedback provided by the hardware, so the user can for instance run a given task a thousand times, measure the global consumption for that series of tasks, divide it by a thousand, repeat for varying kinds of tasks and task sizes, and eventually feed StarPU with these manual measurements through starpu_perfmodel_update_history(). For instance, for CUDA devices, nvidia-smi -q -d POWER
can be used to get the current consumption in Watt. Multiplying that value by the average duration of a single task gives the consumption of the task in Joules, which can be given to starpu_perfmodel_update_history().
In some cases, one may want to force some scheduling, for instance force a given set of tasks to GPU0, another set to GPU1, etc. while letting some other tasks be scheduled on any other device. This can indeed be useful to guide StarPU into some work distribution, while still letting some degree of dynamism. For instance, to force execution of a task on CUDA0:
Note however that using scheduling contexts while statically scheduling tasks on workers could be tricky. Be careful to schedule the tasks exactly on the workers of the corresponding contexts, otherwise the workers' corresponding scheduling structures may not be allocated or the execution of the application may deadlock. Moreover, the hypervisor should not be used when statically scheduling tasks.
A quick view of how many tasks each worker has executed can be obtained by setting 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 export STARPU_BUS_STATS=1
.
More detailed profiling information can be enabled by using export STARPU_PROFILING=1
or by calling starpu_profiling_status_set() from the source code. Statistics on the execution can then be obtained by using export STARPU_BUS_STATS=1
and export STARPU_WORKER_STATS=1
. More details on performance feedback are provided by the next chapter.
It may happen that for some reason, StarPU does not make progress for a long period of time. Reason are sometimes due to contention inside StarPU, but sometimes this is due to external reasons, such as stuck MPI driver, or CUDA driver, etc.
export STARPU_WATCHDOG_TIMEOUT=10000
allows to make StarPU print an error message whenever StarPU does not terminate any task for 10ms. In addition to that,
export STARPU_WATCHDOG_CRASH=1
raises SIGABRT in that condition, thus allowing to catch the situation in gdb. It can also be useful to type "handle SIGABRT nopass" in gdb to be able to let the process continue, after inspecting the state of the process.
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 instead of the default stream, which synchronizes all operations of the GPU. StarPU provides one by the use of starpu_cuda_get_local_stream() which can be used by all CUDA codelet operations to avoid this issue. For instance:
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.
To get an idea of what is happening, a lot of performance feedback is available, detailed in the next 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 Simgrid in order to simulate execution on an arbitrary platform.
There are a few technical details which need to be handled for an application to be simulated through Simgrid.
If the application uses gettimeofday
to make its performance measurements, the real time will be used, which will be bogus. To get the simulated time, it has to use starpu_timing_now() which returns the virtual timestamp in us.
For some technical reason, the application's .c file which contains main() has to be recompiled with starpu.h, which in the simgrid case will #define main() into starpu_main(), and it is libstarpu which will provide the real main() and call the application's main().
To be able to test with crazy data sizes, one may want to only allocate application data if STARPU_SIMGRID is not defined. Passing a NULL pointer to starpu_data_register functions is fine, data will never be read/written to by StarPU in Simgrid mode anyway.
To be able to run the application with e.g. CUDA simulation on a system which does not have CUDA installed, one can fill the cuda_funcs with (void*)1, to express that there is a CUDA implementation, even if one does not actually provide it. StarPU will never actually run it in Simgrid mode anyway.
The idea is to first compile StarPU normally, and run the application, so as to automatically benchmark the bus and the codelets.
$ ./configure && make $ STARPU_SCHED=dmda ./examples/matvecmult/matvecmult [starpu][_starpu_load_history_based_model] Warning: model matvecmult is not calibrated, forcing calibration for this run. Use the STARPU_CALIBRATE environment variable to control this. $ ... $ STARPU_SCHED=dmda ./examples/matvecmult/matvecmult TEST PASSED
Note that we force to use the scheduler dmda
to generate performance models for the application. The application may need to be run several times before the model is calibrated.
Then, recompile StarPU, passing --enable-simgrid to ./configure
.
$ ./configure --enable-simgrid
To specify the location of SimGrid, you can either set the environment variables SIMGRID_CFLAGS and SIMGRID_LIBS, or use the configure options --with-simgrid-dir, --with-simgrid-include-dir and --with-simgrid-lib-dir, for example
$ ./configure --with-simgrid-dir=/opt/local/simgrid
You can then re-run the application.
$ make $ STARPU_SCHED=dmda ./examples/matvecmult/matvecmult TEST FAILED !!!
It is normal that the test fails: since the computation are not actually done (that is the whole point of simgrid), the result is wrong, of course.
If the performance model is not calibrated enough, the following error message will be displayed
$ STARPU_SCHED=dmda ./examples/matvecmult/matvecmult [starpu][_starpu_load_history_based_model] Warning: model matvecmult is not calibrated, forcing calibration for this run. Use the STARPU_CALIBRATE environment variable to control this. [starpu][_starpu_simgrid_execute_job][assert failure] Codelet matvecmult does not have a perfmodel, or is not calibrated enough
The number of devices can be chosen as usual with STARPU_NCPU, STARPU_NCUDA, and STARPU_NOPENCL, and the amount of GPU memory with STARPU_LIMIT_CUDA_MEM, STARPU_LIMIT_CUDA_devid_MEM, STARPU_LIMIT_OPENCL_MEM, and STARPU_LIMIT_OPENCL_devid_MEM.
The simgrid support even permits to perform simulations on another machine, your desktop, typically. To achieve this, one still needs to perform the Calibration step on the actual machine to be simulated, then copy them to your desktop machine (the $STARPU_HOME/.starpu
directory). One can then perform the Simulation step on the desktop machine, by setting the environment variable STARPU_HOSTNAME to the name of the actual machine, to make StarPU use the performance models of the simulated machine even on the desktop machine.
If the desktop machine does not have CUDA or OpenCL, StarPU is still able to use simgrid to simulate execution with CUDA/OpenCL devices, but the application source code will probably disable the CUDA and OpenCL codelets in thatcd sc case. Since during simgrid execution, the functions of the codelet are actually not called, one can use dummy functions such as the following to still permit CUDA or OpenCL execution: