StarPU Handbook
|
In situations where multiple parallel software elements have to coexist within the same application, uncoordinated accesses to computing units may lead such parallel software elements to collide and interfere. The purpose of the Interoperability routines of StarPU, implemented along the definition of the Resource Management APIs of Project H2020 INTERTWinE, is to enable StarPU to coexist with other parallel software elements without resulting in computing core oversubscription or undersubscription. These routines allow the programmer to dynamically control the computing resources allocated to StarPU, to add or remove processor cores and/or accelerator devices from the pool of resources used by StarPU's workers to execute tasks. They also allow multiple libraries and applicative codes using StarPU simultaneously to select distinct sets of resources independently. Internally, the Interoperability Support is built on top of Scheduling Contexts (see Scheduling Contexts).
The starpurm
module is a library built on top of the starpu
library. It exposes a series of routines prefixed with starpurm_
defining the resource management API.
All functions are defined in Interoperability Support.
The starpurm
module must be linked explicitly with the applicative executable using it. Example Makefiles in the starpurm/dev/
subdirectories show how to do so. If the pkg-config
command is available and the PKG_CONFIG_PATH
environment variable is properly positioned, the proper settings may be obtained with the following Makefile
snippet:
The starpurm
module is initialized with a call to starpurm_initialize() and must be finalized with a call to starpurm_shutdown(). The starpurm
module supports CPU cores as well as devices. An integer ID is assigned to each supported device type. The ID assigned to a given device type can be queried with the starpurm_get_device_type_id() routine, which currently expects one of the following strings as argument and returns the corresponding ID:
"cpu"
"opencl"
"cuda"
"mic"
The cpu
pseudo device type is defined for convenience and designates CPU cores. The number of units of each type available for computation can be obtained with a call to starpu_get_nb_devices_by_type().
Each CPU core unit available for computation is designated by its rank among the StarPU CPU worker threads and by its own CPUSET bit. Each non-CPU device unit can be designated both by its rank number in the type, and by the CPUSET bit corresponding to its StarPU device worker thread. The CPUSET of a computing unit or its associated worker can be obtained from its type ID and rank with starpurm_get_device_worker_cpuset(), which returns the corresponding HWLOC CPUSET.
The starpurm
module assumes a default, global context, manipulated through a series of routines allowing to assign and withdraw computing units from the main StarPU context. Assigning CPU cores can be done with starpurm_assign_cpu_to_starpu() and starpurm_assign_cpu_mask_to_starpu(), and assigning device units can be done with starpurm_assign_device_to_starpu() and starpurm_assign_device_mask_to_starpu(). Conversely, withdrawing CPU cores can be done with starpurm_withdraw_cpu_from_starpu() and starpurm_withdraw_cpu_mask_from_starpu(=, and withdrawing device units can be done with starpurm_withdraw_device_from_starpu() and starpurm_withdraw_device_mask_from_starpu(). These routine should typically be used to control resource usage for the main applicative code.
Besides the default, global context, starpurm
can create temporary contexts and launch the computation of kernels confined to these temporary contexts. The routine starpurm_spawn_kernel_on_cpus() can be used to do so: it allocates a temporary context and spawns a kernel within this context. The temporary context is subsequently freed upon completion of the kernel. The temporary context is set as the default context for the kernel throughout its lifespan. This routine should typically be used to control resource usage for a parallel kernel handled by an external library built on StarPU. Internally, it relies on the use of starpu_sched_ctx_set_context() to set the temporary context as default context for the parallel kernel, and then restore the main context upon completion. Note: the maximum number of temporary contexts allocated concurrently at any time should not exceed STARPU_NMAX_SCHED_CTXS-2, otherwise, the call to starpurm_spawn_kernel_on_cpus() may block until a temporary context becomes available. The routine starpurm_spawn_kernel_on_cpus() returns upon the completion of the parallel kernel. An asynchronous variant is available with the routine starpurm_spawn_kernel_on_cpus_callback(). This variant returns immediately, however it accepts a callback function, which is subsequently called to notify the calling code about the completion of the parallel kernel.