|
@@ -8,50 +8,104 @@
|
|
|
#include <hwloc.h>
|
|
|
#endif
|
|
|
|
|
|
+/* struct _starpu_sched_node are scheduler modules, a scheduler is a tree-like
|
|
|
+ * structure of them, some parts of scheduler can be shared by several contexes
|
|
|
+ * to perform some local optimisations, so, for all nodes, a list of father is
|
|
|
+ * defined indexed by sched_ctx_id
|
|
|
+ *
|
|
|
+ * they embed there specialised method in a pseudo object-style, so calls are like node->push_task(node,task)
|
|
|
+ *
|
|
|
+ */
|
|
|
struct _starpu_sched_node
|
|
|
{
|
|
|
+ /* node->push_task(node, task)
|
|
|
+ * this function is called to push a task on node subtree, this can either
|
|
|
+ * perform a recursive call on a child or store the task in the node, then
|
|
|
+ * it will be returned by a further pop_task call
|
|
|
+ *
|
|
|
+ * the caller must ensure that node is able to execute task
|
|
|
+ */
|
|
|
int (*push_task)(struct _starpu_sched_node *,
|
|
|
struct starpu_task *);
|
|
|
+ /* this function is called by workers to get a task on them fathers
|
|
|
+ * this function should first return a localy stored task or perform
|
|
|
+ * a recursive call on father
|
|
|
+ *
|
|
|
+ * a default implementation simply do a recursive call on father
|
|
|
+ */
|
|
|
struct starpu_task * (*pop_task)(struct _starpu_sched_node *,
|
|
|
unsigned sched_ctx_id);
|
|
|
|
|
|
+ /* this function notify underlying worker that a task as been pushed
|
|
|
+ * and would be returned by a pop_task call
|
|
|
+ * it should be called each time a node localy store a task
|
|
|
+ *
|
|
|
+ * default implementation simply perform a recursive call on childrens
|
|
|
+ * this function can be called by a worker as it doesn't try to wake up himself
|
|
|
+ */
|
|
|
void (*available)(struct _starpu_sched_node *);
|
|
|
|
|
|
- /* this function is an heuristic that compute load of subtree */
|
|
|
+ /* this function is an heuristic that compute load of subtree, basicaly
|
|
|
+ * it compute
|
|
|
+ * estimated_load(node) = sum(estimated_load(node_childs)) +
|
|
|
+ * nb_local_tasks / average(relative_speedup(underlying_worker))
|
|
|
+ */
|
|
|
double (*estimated_load)(struct _starpu_sched_node * node);
|
|
|
|
|
|
+ /* this function return a struct _starpu_task_execute_preds defined lower
|
|
|
+ * wich basicaly give predictions for a task execution a call on
|
|
|
+ * homogeneous (with all workers of the same arch) node is optimised
|
|
|
+ */
|
|
|
struct _starpu_task_execute_preds (*estimated_execute_preds)(struct _starpu_sched_node * node,
|
|
|
struct starpu_task * task);
|
|
|
-
|
|
|
+ /* the numbers of node's childs
|
|
|
+ */
|
|
|
int nchilds;
|
|
|
+ /* the vector of node's childs
|
|
|
+ */
|
|
|
struct _starpu_sched_node ** childs;
|
|
|
-
|
|
|
- //the set of workers in the node's subtree
|
|
|
- struct _starpu_bitmap * workers;
|
|
|
- //is_homogeneous is 0 iff workers in the node's subtree are heterogeneous,
|
|
|
- //this field is set and updated automaticaly, you shouldn't write on it
|
|
|
- int is_homogeneous;
|
|
|
-
|
|
|
- void * data;
|
|
|
/* may be shared by several contexts
|
|
|
* so we need several fathers
|
|
|
*/
|
|
|
struct _starpu_sched_node * fathers[STARPU_NMAX_SCHED_CTXS];
|
|
|
+ /* the set of workers in the node's subtree
|
|
|
+ */
|
|
|
+ struct _starpu_bitmap * workers;
|
|
|
+ /* is_homogeneous is 0 iff workers in the node's subtree are heterogeneous,
|
|
|
+ * this field is set and updated automaticaly, you shouldn't write on it
|
|
|
+ */
|
|
|
+ int is_homogeneous;
|
|
|
+ /* node's private data, no restriction on use
|
|
|
+ */
|
|
|
+ void * data;
|
|
|
|
|
|
- /* this function is called after all childs has been set
|
|
|
+ /* this function is called after all childs has been set, and the
|
|
|
+ * workers member was filled, can be used to init data, or anything you want
|
|
|
*/
|
|
|
void (*init_data)(struct _starpu_sched_node *);
|
|
|
/* this function is called to free data allocated by init_data
|
|
|
+ * just before the call of _starpu_sched_node_destroy(node)
|
|
|
*/
|
|
|
void (*deinit_data)(struct _starpu_sched_node *);
|
|
|
|
|
|
#ifdef STARPU_HAVE_HWLOC
|
|
|
+ /* in case of a hierarchical scheduler, this is set to the part of
|
|
|
+ * topology that is binded to this node, eg: a numa node for a ws
|
|
|
+ * node that would balance load between underlying sockets
|
|
|
+ */
|
|
|
hwloc_obj_t obj;
|
|
|
#endif
|
|
|
};
|
|
|
|
|
|
+/* this structure is only returned by estimated_execute_preds and give
|
|
|
+ * predictions on task computations
|
|
|
+ */
|
|
|
struct _starpu_task_execute_preds
|
|
|
{
|
|
|
+ /* if several value are possible for state member,
|
|
|
+ * in order of priority :
|
|
|
+ * CALIBRATING, PERF_MODEL, NO_PERF_MODEL, CANNOT_EXECUTE
|
|
|
+ */
|
|
|
enum {CANNOT_EXECUTE = 0, CALIBRATING , NO_PERF_MODEL, PERF_MODEL} state;
|
|
|
|
|
|
/* archtype and nimpl is set to
|
|
@@ -74,7 +128,10 @@ struct _starpu_sched_tree
|
|
|
{
|
|
|
struct _starpu_sched_node * root;
|
|
|
struct _starpu_bitmap * workers;
|
|
|
- //this lock is used to protect the scheduler during modifications of his structure
|
|
|
+ /* this lock is used to protect the scheduler,
|
|
|
+ * it is taken in read mode pushing a task
|
|
|
+ * and in write mode for adding or removing workers
|
|
|
+ */
|
|
|
starpu_pthread_rwlock_t lock;
|
|
|
};
|
|
|
|