StarPU Handbook
|
Data management facilities provided by StarPU. We show how to use existing data interfaces in Data Interfaces, but developers can design their own data interfaces if required. More...
Typedefs | |
typedef struct _starpu_data_state * | starpu_data_handle_t |
typedef struct starpu_arbiter * | starpu_arbiter_t |
Enumerations | |
enum | starpu_data_access_mode { STARPU_NONE , STARPU_R , STARPU_W , STARPU_RW , STARPU_SCRATCH , STARPU_REDUX , STARPU_COMMUTE , STARPU_SSEND , STARPU_LOCALITY , STARPU_NOPLAN , STARPU_ACCESS_MODE_MAX } |
Access registered data from the application | |
int | starpu_data_acquire (starpu_data_handle_t handle, enum starpu_data_access_mode mode) |
int | starpu_data_acquire_on_node (starpu_data_handle_t handle, int node, enum starpu_data_access_mode mode) |
int | starpu_data_acquire_cb (starpu_data_handle_t handle, enum starpu_data_access_mode mode, void(*callback)(void *), void *arg) |
int | starpu_data_acquire_on_node_cb (starpu_data_handle_t handle, int node, enum starpu_data_access_mode mode, void(*callback)(void *), void *arg) |
int | starpu_data_acquire_cb_sequential_consistency (starpu_data_handle_t handle, enum starpu_data_access_mode mode, void(*callback)(void *), void *arg, int sequential_consistency) |
int | starpu_data_acquire_on_node_cb_sequential_consistency (starpu_data_handle_t handle, int node, enum starpu_data_access_mode mode, void(*callback)(void *), void *arg, int sequential_consistency) |
int | starpu_data_acquire_on_node_cb_sequential_consistency_quick (starpu_data_handle_t handle, int node, enum starpu_data_access_mode mode, void(*callback)(void *), void *arg, int sequential_consistency, int quick) |
int | starpu_data_acquire_on_node_cb_sequential_consistency_sync_jobids (starpu_data_handle_t handle, int node, enum starpu_data_access_mode mode, void(*callback)(void *), void *arg, int sequential_consistency, int quick, long *pre_sync_jobid, long *post_sync_jobid) |
int | starpu_data_acquire_try (starpu_data_handle_t handle, enum starpu_data_access_mode mode) |
int | starpu_data_acquire_on_node_try (starpu_data_handle_t handle, int node, enum starpu_data_access_mode mode) |
void | starpu_data_release (starpu_data_handle_t handle) |
void | starpu_data_release_on_node (starpu_data_handle_t handle, int node) |
#define | STARPU_ACQUIRE_NO_NODE |
#define | STARPU_ACQUIRE_NO_NODE_LOCK_ALL |
#define | STARPU_DATA_ACQUIRE_CB(handle, mode, code) |
Implicit Data Dependencies | |
In this section, we describe how StarPU makes it possible to insert implicit task dependencies in order to enforce sequential data consistency. When this data consistency is enabled on a specific data handle, any data access will appear as sequentially consistent from the application. For instance, if the application submits two tasks that access the same piece of data in read-only mode, and then a third task that access it in write mode, dependencies will be added between the two first tasks and the third one. Implicit data dependencies are also inserted in the case of data accesses from the application. | |
void | starpu_data_set_sequential_consistency_flag (starpu_data_handle_t handle, unsigned flag) |
unsigned | starpu_data_get_sequential_consistency_flag (starpu_data_handle_t handle) |
unsigned | starpu_data_get_default_sequential_consistency_flag (void) |
void | starpu_data_set_default_sequential_consistency_flag (unsigned flag) |
Data management facilities provided by StarPU. We show how to use existing data interfaces in Data Interfaces, but developers can design their own data interfaces if required.
#define STARPU_ACQUIRE_NO_NODE |
This macro can be used to acquire data, but not require it to be available on a given node, only enforce R/W dependencies. This can for instance be used to wait for tasks which produce the data, but without requesting a fetch to the main memory.
#define STARPU_ACQUIRE_NO_NODE_LOCK_ALL |
Similar to STARPU_ACQUIRE_NO_NODE, but will lock the data on all nodes, preventing them from being evicted for instance. This is mostly useful inside StarPU only.
#define STARPU_DATA_ACQUIRE_CB | ( | handle, | |
mode, | |||
code | |||
) |
STARPU_DATA_ACQUIRE_CB() is the same as starpu_data_acquire_cb(), except that the code to be executed in a callback is directly provided as a macro parameter, and the data handle
is automatically released after it. This permits to easily execute code which depends on the value of some registered data. This is non-blocking too and may be called from task callbacks.
typedef struct _starpu_data_state* starpu_data_handle_t |
StarPU uses starpu_data_handle_t as an opaque handle to manage a piece of data. Once a piece of data has been registered to StarPU, it is associated to a starpu_data_handle_t which keeps track of the state of the piece of data over the entire machine, so that we can maintain data consistency and locate data replicates for instance.
typedef struct starpu_arbiter* starpu_arbiter_t |
This is an arbiter, which implements an advanced but centralized management of concurrent data accesses, see Concurrent Data Accesses for the details.
Describe a StarPU data access mode
Note: when adding a flag here, update _starpu_detect_implicit_data_deps_with_handle
Note: other STARPU_* values in include/starpu_task_util.h
Enumerator | |
---|---|
STARPU_NONE | todo |
STARPU_R | read-only mode |
STARPU_W | write-only mode |
STARPU_RW | |
STARPU_SCRATCH | A temporary buffer is allocated for the task, but StarPU does not enforce data consistency—i.e. each device has its own buffer, independently from each other (even for CPUs), and no data transfer is ever performed. This is useful for temporary variables to avoid allocating/freeing buffers inside each task. Currently, no behavior is defined concerning the relation with the STARPU_R and STARPU_W modes and the value provided at registration — i.e., the value of the scratch buffer is undefined at entry of the codelet function. It is being considered for future extensions at least to define the initial value. For now, data to be used in STARPU_SCRATCH mode should be registered with node -1 and a |
STARPU_REDUX | Reduction mode. StarPU will allocate on the fly a per-worker buffer, so that various tasks that access the same data in STARPU_REDUX mode can execute in parallel. When a task accesses the data without STARPU_REDUX, StarPU will automatically reduce the different contributions. Codelets contributing to these reductions with STARPU_REDUX must be registered with STARPU_RW | STARPU_COMMUTE access modes. See Data Reduction for more details. |
STARPU_COMMUTE | STARPU_COMMUTE can be passed along STARPU_W or STARPU_RW to express that StarPU can let tasks commute, which is useful e.g. when bringing a contribution into some data, which can be done in any order (but still require sequential consistency against reads or non-commutative writes). |
STARPU_SSEND | used in starpu_mpi_insert_task() to specify the data has to be sent using a synchronous and non-blocking mode (see starpu_mpi_issend()) |
STARPU_LOCALITY | used to tell the scheduler which data is the most important for the task, and should thus be used to try to group tasks on the same core or cache, etc. For now only the ws and lws schedulers take this flag into account, and only when rebuild with |
STARPU_NOPLAN | Disable automatic submission of asynchronous partitioning/unpartitioning |
STARPU_ACCESS_MODE_MAX | The purpose of STARPU_ACCESS_MODE_MAX is to be the maximum of this enum. |
void starpu_data_set_name | ( | starpu_data_handle_t | handle, |
const char * | name | ||
) |
Set the name of the data, to be shown in various profiling tools.
void starpu_data_set_coordinates_array | ( | starpu_data_handle_t | handle, |
unsigned | dimensions, | ||
int | dims[] | ||
) |
Set the coordinates of the data, to be shown in various profiling tools. dimensions
is the size of the dims
array. This can be for instance the tile coordinates within a big matrix.
void starpu_data_set_coordinates | ( | starpu_data_handle_t | handle, |
unsigned | dimensions, | ||
... | |||
) |
Set the coordinates of the data, to be shown in various profiling tools. dimensions
is the number of subsequent int
parameters. This can be for instance the tile coordinates within a big matrix.
unsigned starpu_data_get_coordinates_array | ( | starpu_data_handle_t | handle, |
unsigned | dimensions, | ||
int | dims[] | ||
) |
Get the coordinates of the data, as set by a previous call to starpu_data_set_coordinates_array() or starpu_data_set_coordinates() dimensions
is the size of the dims
array. This returns the actual number of returned coordinates.
void starpu_data_unregister | ( | starpu_data_handle_t | handle | ) |
Unregister a data handle
from StarPU. If the data was automatically allocated by StarPU because the home node was -1, all automatically allocated buffers are freed. Otherwise, a valid copy of the data is put back into the home node in the buffer that was initially registered. Using a data handle that has been unregistered from StarPU results in an undefined behaviour. In case we do not need to update the value of the data in the home node, we can use the function starpu_data_unregister_no_coherency() instead.
void starpu_data_unregister_no_coherency | ( | starpu_data_handle_t | handle | ) |
Similar to starpu_data_unregister(), except that StarPU does not put back a valid copy into the home node, in the buffer that was initially registered.
void starpu_data_unregister_submit | ( | starpu_data_handle_t | handle | ) |
Destroy the data handle
once it is no longer needed by any submitted task. No coherency is provided.
This is not safe to call starpu_data_unregister_submit() on a handle that comes from the registration of a non-NULL application home buffer, since the moment when the unregistration will happen is unknown to the application. Only calling starpu_shutdown() allows to be sure that the data was really unregistered.
void starpu_data_invalidate | ( | starpu_data_handle_t | handle | ) |
void starpu_data_invalidate_submit | ( | starpu_data_handle_t | handle | ) |
Submit invalidation of the data handle
after completion of previously submitted tasks.
void starpu_data_advise_as_important | ( | starpu_data_handle_t | handle, |
unsigned | is_important | ||
) |
Specify that the data handle
can be discarded without impacting the application.
int starpu_data_acquire | ( | starpu_data_handle_t | handle, |
enum starpu_data_access_mode | mode | ||
) |
The application must call this function prior to accessing registered data from main memory outside tasks. StarPU ensures that the application will get an up-to-date copy of handle
in main memory located where the data was originally registered, and that all concurrent accesses (e.g. from tasks) will be consistent with the access mode specified with mode
. starpu_data_release() must be called once the application no longer needs to access the piece of data. Note that implicit data dependencies are also enforced by starpu_data_acquire(), i.e. starpu_data_acquire() will wait for all tasks scheduled to work on the data, unless they have been disabled explictly by calling starpu_data_set_default_sequential_consistency_flag() or starpu_data_set_sequential_consistency_flag(). starpu_data_acquire() is a blocking call, so that it cannot be called from tasks or from their callbacks (in that case, starpu_data_acquire() returns -EDEADLK
). Upon successful completion, this function returns 0.
int starpu_data_acquire_on_node | ( | starpu_data_handle_t | handle, |
int | node, | ||
enum starpu_data_access_mode | mode | ||
) |
Similar to starpu_data_acquire(), except that the data will be available on the given memory node instead of main memory. STARPU_ACQUIRE_NO_NODE and STARPU_ACQUIRE_NO_NODE_LOCK_ALL can be used instead of an explicit node number.
int starpu_data_acquire_cb | ( | starpu_data_handle_t | handle, |
enum starpu_data_access_mode | mode, | ||
void(*)(void *) | callback, | ||
void * | arg | ||
) |
Asynchronous equivalent of starpu_data_acquire(). When the data specified in handle
is available in the access mode
, the callback
function is executed. The application may access the requested data during the execution of callback
. The callback
function must call starpu_data_release() once the application no longer needs to access the piece of data. Note that implicit data dependencies are also enforced by starpu_data_acquire_cb() in case they are not disabled. Contrary to starpu_data_acquire(), this function is non-blocking and may be called from task callbacks. Upon successful completion, this function returns 0.
int starpu_data_acquire_on_node_cb | ( | starpu_data_handle_t | handle, |
int | node, | ||
enum starpu_data_access_mode | mode, | ||
void(*)(void *) | callback, | ||
void * | arg | ||
) |
Similar to starpu_data_acquire_cb(), except that the data will be available on the given memory node instead of main memory. STARPU_ACQUIRE_NO_NODE and STARPU_ACQUIRE_NO_NODE_LOCK_ALL can be used instead of an explicit node number.
int starpu_data_acquire_cb_sequential_consistency | ( | starpu_data_handle_t | handle, |
enum starpu_data_access_mode | mode, | ||
void(*)(void *) | callback, | ||
void * | arg, | ||
int | sequential_consistency | ||
) |
Similar to starpu_data_acquire_cb() with the possibility of enabling or disabling data dependencies. When the data specified in handle
is available in the access mode
, the callback
function is executed. The application may access the requested data during the execution of this callback
. The callback
function must call starpu_data_release() once the application no longer needs to access the piece of data. Note that implicit data dependencies are also enforced by starpu_data_acquire_cb_sequential_consistency() in case they are not disabled specifically for the given handle
or by the parameter sequential_consistency
. Similarly to starpu_data_acquire_cb(), this function is non-blocking and may be called from task callbacks. Upon successful completion, this function returns 0.
int starpu_data_acquire_on_node_cb_sequential_consistency | ( | starpu_data_handle_t | handle, |
int | node, | ||
enum starpu_data_access_mode | mode, | ||
void(*)(void *) | callback, | ||
void * | arg, | ||
int | sequential_consistency | ||
) |
Similar to starpu_data_acquire_cb_sequential_consistency(), except that the data will be available on the given memory node instead of main memory. STARPU_ACQUIRE_NO_NODE and STARPU_ACQUIRE_NO_NODE_LOCK_ALL can be used instead of an explicit node number.
int starpu_data_acquire_on_node_cb_sequential_consistency_sync_jobids | ( | starpu_data_handle_t | handle, |
int | node, | ||
enum starpu_data_access_mode | mode, | ||
void(*)(void *) | callback, | ||
void * | arg, | ||
int | sequential_consistency, | ||
int | quick, | ||
long * | pre_sync_jobid, | ||
long * | post_sync_jobid | ||
) |
Similar to starpu_data_acquire_on_node_cb_sequential_consistency(), except that the pre_sync_jobid and post_sync_jobid parameters can be used to retrieve the jobid of the synchronization tasks. pre_sync_jobid happens just before the acquisition, and post_sync_jobid happens just after the release.
int starpu_data_acquire_try | ( | starpu_data_handle_t | handle, |
enum starpu_data_access_mode | mode | ||
) |
The application can call this function instead of starpu_data_acquire() so as to acquire the data like starpu_data_acquire(), but only if all previously-submitted tasks have completed, in which case starpu_data_acquire_try() returns 0. StarPU will have ensured that the application will get an up-to-date copy of handle
in main memory located where the data was originally registered. starpu_data_release() must be called once the application no longer needs to access the piece of data.
int starpu_data_acquire_on_node_try | ( | starpu_data_handle_t | handle, |
int | node, | ||
enum starpu_data_access_mode | mode | ||
) |
Similar to starpu_data_acquire_try(), except that the data will be available on the given memory node instead of main memory. STARPU_ACQUIRE_NO_NODE and STARPU_ACQUIRE_NO_NODE_LOCK_ALL can be used instead of an explicit node number.
void starpu_data_release | ( | starpu_data_handle_t | handle | ) |
Release the piece of data acquired by the application either by starpu_data_acquire() or by starpu_data_acquire_cb().
void starpu_data_release_on_node | ( | starpu_data_handle_t | handle, |
int | node | ||
) |
Similar to starpu_data_release(), except that the data will be available on the given memory node
instead of main memory. The node
parameter must be exactly the same as the corresponding starpu_data_acquire_on_node*
call.
starpu_arbiter_t starpu_arbiter_create | ( | void | ) |
Create a data access arbiter, see Concurrent Data Accesses for the details
void starpu_data_assign_arbiter | ( | starpu_data_handle_t | handle, |
starpu_arbiter_t | arbiter | ||
) |
Make access to handle
managed by arbiter
void starpu_arbiter_destroy | ( | starpu_arbiter_t | arbiter | ) |
Destroy the arbiter
. This must only be called after all data assigned to it have been unregistered.
int starpu_data_request_allocation | ( | starpu_data_handle_t | handle, |
unsigned | node | ||
) |
Explicitly ask StarPU to allocate room for a piece of data on the specified memory node
.
int starpu_data_fetch_on_node | ( | starpu_data_handle_t | handle, |
unsigned | node, | ||
unsigned | async | ||
) |
Issue a fetch request for the data handle
to node
, i.e. requests that the data be replicated to the given node as soon as possible, so that it is available there for tasks. If async
is 0, the call will block until the transfer is achieved, else the call will return immediately, after having just queued the request. In the latter case, the request will asynchronously wait for the completion of any task writing on the data.
int starpu_data_prefetch_on_node | ( | starpu_data_handle_t | handle, |
unsigned | node, | ||
unsigned | async | ||
) |
Issue a prefetch request for the data handle
to node
, i.e. requests that the data be replicated to node
when there is room for it, so that it is available there for tasks. If async
is 0, the call will block until the transfer is achieved, else the call will return immediately, after having just queued the request. In the latter case, the request will asynchronously wait for the completion of any task writing on the data.
int starpu_data_idle_prefetch_on_node | ( | starpu_data_handle_t | handle, |
unsigned | node, | ||
unsigned | async | ||
) |
Issue an idle prefetch request for the data handle
to node
, i.e. requests that the data be replicated to node
, so that it is available there for tasks, but only when the bus is really idle. If async
is 0, the call will block until the transfer is achieved, else the call will return immediately, after having just queued the request. In the latter case, the request will asynchronously wait for the completion of any task writing on the data.
unsigned starpu_data_is_on_node | ( | starpu_data_handle_t | handle, |
unsigned | node | ||
) |
Check whether a valid copy of handle
is currently available on memory node node
.
void starpu_data_wont_use | ( | starpu_data_handle_t | handle | ) |
Advise StarPU that handle
will not be used in the close future, and is thus a good candidate for eviction from GPUs. StarPU will thus write its value back to its home node when the bus is idle, and select this data in priority for eviction when memory gets low.
void starpu_data_set_wt_mask | ( | starpu_data_handle_t | handle, |
uint32_t | wt_mask | ||
) |
Set the write-through mask of the data handle
(and its children), i.e. a bitmask of nodes where the data should be always replicated after modification. It also prevents the data from being evicted from these nodes when memory gets scarse. When the data is modified, it is automatically transfered into those memory nodes. For instance a 1<<0
write-through mask means that the CUDA workers will commit their changes in main memory (node 0).
void starpu_data_set_sequential_consistency_flag | ( | starpu_data_handle_t | handle, |
unsigned | flag | ||
) |
Set the data consistency mode associated to a data handle. The consistency mode set using this function has the priority over the default mode which can be set with starpu_data_set_default_sequential_consistency_flag().
unsigned starpu_data_get_sequential_consistency_flag | ( | starpu_data_handle_t | handle | ) |
Get the data consistency mode associated to the data handle handle
unsigned starpu_data_get_default_sequential_consistency_flag | ( | void | ) |
Return the default sequential consistency flag
void starpu_data_set_default_sequential_consistency_flag | ( | unsigned | flag | ) |
Set the default sequential consistency flag. If a non-zero value is passed, a sequential data consistency will be enforced for all handles registered after this function call, otherwise it is disabled. By default, StarPU enables sequential data consistency. It is also possible to select the data consistency mode of a specific data handle with the function starpu_data_set_sequential_consistency_flag().
void starpu_data_set_ooc_flag | ( | starpu_data_handle_t | handle, |
unsigned | flag | ||
) |
Set whether this data should be elligible to be evicted to disk storage (1) or not (0). The default is 1.
unsigned starpu_data_get_ooc_flag | ( | starpu_data_handle_t | handle | ) |
Get whether this data was set to be elligible to be evicted to disk storage (1) or not (0).
void starpu_data_query_status | ( | starpu_data_handle_t | handle, |
int | memory_node, | ||
int * | is_allocated, | ||
int * | is_valid, | ||
int * | is_requested | ||
) |
Query the status of handle
on the specified memory_node
.
void starpu_data_set_reduction_methods | ( | starpu_data_handle_t | handle, |
struct starpu_codelet * | redux_cl, | ||
struct starpu_codelet * | init_cl | ||
) |
Set the codelets to be used for handle
when it is accessed in the mode STARPU_REDUX. Per-worker buffers will be initialized with the codelet init_cl
, and reduction between per-worker buffers will be done with the codelet redux_cl
.
void starpu_data_set_user_data | ( | starpu_data_handle_t | handle, |
void * | user_data | ||
) |
Set the field user_data
for the handle
to user_data
. It can then be retrieved with starpu_data_get_user_data(). user_data
can be any application-defined value, for instance a pointer to an object-oriented container for the data.
void* starpu_data_get_user_data | ( | starpu_data_handle_t | handle | ) |
Retrieve the field user_data
previously set for the handle
.