Zoltan2
|
#include <Zoltan2_TaskMapping.hpp>
Public Member Functions | |
void | getProcTask (part_t *&proc_to_task_xadj_, part_t *&proc_to_task_adj_) |
virtual void | map (const RCP< MappingSolution< Adapter > > &mappingsoln) |
Mapping method. More... | |
virtual | ~CoordinateTaskMapper () |
void | create_local_task_to_rank (const lno_t num_local_coords, const part_t *local_coord_parts, const ArrayRCP< part_t > task_to_proc_) |
CoordinateTaskMapper (const Teuchos::RCP< const Teuchos::Comm< int > > comm_, const Teuchos::RCP< const MachineRepresentation< pcoord_t, part_t > > machine_, const Teuchos::RCP< const Adapter > input_adapter_, const Teuchos::RCP< const Zoltan2::PartitioningSolution< Adapter > > soln_, const Teuchos::RCP< const Environment > envConst, bool is_input_adapter_distributed=true, int num_ranks_per_node=1, bool divide_to_prime_first=false, bool reduce_best_mapping=true) | |
Constructor. When this constructor is called, in order to calculate the communication metric, the task adjacency graph is created based on the coordinate model input and partitioning of it. If the communication graph is already calculated, use the other constructors. More... | |
CoordinateTaskMapper (const Teuchos::RCP< const Teuchos::Comm< int > > comm_, const Teuchos::RCP< const MachineRepresentation< pcoord_t, part_t > > machine_, const Teuchos::RCP< const Adapter > input_adapter_, const part_t num_parts_, const part_t *result_parts, const Teuchos::RCP< const Environment > envConst, bool is_input_adapter_distributed=true, int num_ranks_per_node=1, bool divide_to_prime_first=false, bool reduce_best_mapping=true) | |
Constructor. Instead of Solution we have two parameters, numparts. More... | |
CoordinateTaskMapper (const Environment *env_const_, const Teuchos::Comm< int > *problemComm, int proc_dim, int num_processors, pcoord_t **machine_coords, int task_dim, part_t num_tasks, tcoord_t **task_coords, ArrayRCP< part_t >task_comm_xadj, ArrayRCP< part_t >task_comm_adj, pcoord_t *task_communication_edge_weight_, int recursion_depth, Kokkos::View< part_t *, Kokkos::HostSpace > part_no_array, const part_t *machine_dimensions, int num_ranks_per_node=1, bool divide_to_prime_first=false, bool reduce_best_mapping=true) | |
Constructor The mapping constructor which will also perform the mapping operation. The result mapping can be obtained by –getAssignedProcForTask function: which returns the assigned processor id for the given task –getPartsForProc: which returns the assigned tasks with the number of tasks. More... | |
virtual size_t | getLocalNumberOfParts () const |
Returns the number of parts to be assigned to this process. More... | |
pcoord_t ** | shiftMachineCoordinates (int machine_dim, const part_t *machine_dimensions, bool *machine_extent_wrap_around, part_t numProcs, pcoord_t **mCoords) |
Using the machine dimensions provided, create virtual machine coordinates by assigning the largest gap to be as the wrap around link. More... | |
virtual void | getProcsForPart (part_t taskId, part_t &numProcs, part_t *&procs) const |
getAssignedProcForTask function, returns the assigned tasks with the number of tasks. More... | |
part_t | getAssignedProcForTask (part_t taskId) |
getAssignedProcForTask function, returns the assigned processor id for the given task More... | |
virtual void | getPartsForProc (int procId, part_t &numParts, part_t *&parts) const |
getAssignedProcForTask function, returns the assigned tasks with the number of tasks. More... | |
ArrayView< part_t > | getAssignedTasksForProc (part_t procId) |
![]() | |
PartitionMapping (const Teuchos::RCP< const Teuchos::Comm< int > >comm_, const Teuchos::RCP< const Zoltan2::MachineRepresentation< pcoord_t, part_t > >machine_, const Teuchos::RCP< const Adapter > input_adapter_, const Teuchos::RCP< const Zoltan2::PartitioningSolution< Adapter > >soln_, const Teuchos::RCP< const Environment > envConst_) | |
Constructor Constructor builds the map from parts to ranks. KDDKDD WILL NEED THE SOLUTION FOR INTELLIGENT MAPPING KDDKDD BUT MAY WANT TO SET PART SIZES BASED ON CAPABILITY OF A RANK. KDDKDD SO WHEN SHOULD THE MAP BE CREATED? More... | |
PartitionMapping (const Teuchos::RCP< const Teuchos::Comm< int > >comm_, const Teuchos::RCP< const Zoltan2::MachineRepresentation< pcoord_t, part_t > >machine_, const Teuchos::RCP< const Adapter > input_adapter_, const part_t num_parts_, const part_t *result_parts, const Teuchos::RCP< const Environment > envConst_) | |
PartitionMapping (const Teuchos::RCP< const Teuchos::Comm< int > >comm_, const Teuchos::RCP< const Environment > envConst_) | |
PartitionMapping () | |
PartitionMapping (const Teuchos::RCP< const Environment >envConst_) | |
PartitionMapping (const Teuchos::RCP< const Environment > envConst_, const Teuchos::RCP< const Teuchos::Comm< int > >comm_, const Teuchos::RCP< const MachineRepresentation< pcoord_t, part_t > >machine_) | |
virtual | ~PartitionMapping () |
![]() | |
virtual | ~Algorithm () |
virtual int | localOrder (const RCP< LocalOrderingSolution< lno_t > > &) |
Ordering method. More... | |
virtual int | globalOrder (const RCP< GlobalOrderingSolution< gno_t > > &) |
Ordering method. More... | |
virtual void | color (const RCP< ColoringSolution< Adapter > > &) |
Coloring method. More... | |
virtual void | match () |
Matching method. More... | |
virtual void | partition (const RCP< PartitioningSolution< Adapter > > &) |
Partitioning method. More... | |
virtual void | partitionMatrix (const RCP< MatrixPartitioningSolution< Adapter > > &) |
Matrix Partitioning method. More... | |
virtual bool | isPartitioningTreeBinary () const |
return if algorithm determins tree to be binary More... | |
virtual void | getPartitionTree (part_t, part_t &, std::vector< part_t > &, std::vector< part_t > &, std::vector< part_t > &, std::vector< part_t > &) const |
for partitioning methods, fill arrays with partition tree info More... | |
virtual std::vector < coordinateModelPartBox > & | getPartBoxesView () const |
for partitioning methods, return bounding boxes of the More... | |
virtual part_t | pointAssign (int, scalar_t *) const |
pointAssign method: Available only for some partitioning algorithms More... | |
virtual void | boxAssign (int, scalar_t *, scalar_t *, size_t &, part_t **) const |
boxAssign method: Available only for some partitioning algorithms More... | |
virtual void | getCommunicationGraph (const PartitioningSolution< Adapter > *, ArrayRCP< part_t > &, ArrayRCP< part_t > &) |
returns serial communication graph of a computed partition More... | |
virtual int | getRankForPart (part_t) |
In mapping, returns the rank to which a part is assigned. More... | |
virtual void | getMyPartsView (part_t &, part_t *&) |
In mapping, returns a view of parts assigned to the current rank. More... | |
Protected Member Functions | |
void | doMapping (int myRank, const Teuchos::RCP< const Teuchos::Comm< int > > comm_) |
doMapping function, calls getMapping function of communicationModel object. More... | |
RCP< Comm< int > > | create_subCommunicator () |
creates and returns the subcommunicator for the processor group. More... | |
void | getBestMapping () |
finds the lowest cost mapping and broadcasts solution to everyone. More... | |
void | writeMapping () |
void | writeMapping2 (int myRank) |
Protected Attributes | |
ArrayRCP< part_t > | proc_to_task_xadj |
ArrayRCP< part_t > | proc_to_task_adj |
ArrayRCP< part_t > | task_to_proc |
ArrayRCP< part_t > | local_task_to_rank |
bool | isOwnerofModel |
CoordinateCommunicationModel < pcoord_t, tcoord_t, part_t, node_t > * | proc_task_comm |
part_t | nprocs |
part_t | ntasks |
ArrayRCP< part_t > | task_communication_xadj |
ArrayRCP< part_t > | task_communication_adj |
ArrayRCP< scalar_t > | task_communication_edge_weight |
Additional Inherited Members | |
![]() | |
typedef Adapter::lno_t | lno_t |
typedef Adapter::gno_t | gno_t |
typedef Adapter::scalar_t | scalar_t |
typedef Adapter::part_t | part_t |
![]() | |
const Teuchos::RCP< const Teuchos::Comm< int > > | comm |
const Teuchos::RCP< const Zoltan2::MachineRepresentation < pcoord_t, part_t > > | machine |
const Teuchos::RCP< const Adapter > | input_adapter |
const Teuchos::RCP< const Zoltan2::PartitioningSolution < Adapter > > | soln |
const Teuchos::RCP< const Environment > | env |
const part_t | num_parts |
const part_t * | solution_parts |
Definition at line 1758 of file Zoltan2_TaskMapping.hpp.
|
inlinevirtual |
Definition at line 2226 of file Zoltan2_TaskMapping.hpp.
|
inline |
Constructor. When this constructor is called, in order to calculate the communication metric, the task adjacency graph is created based on the coordinate model input and partitioning of it. If the communication graph is already calculated, use the other constructors.
comm_ | is the communication object. |
machine_ | is the machineRepresentation object. Stores the coordinates of machines. |
model_ | is the input adapter. |
soln_ | is the solution object. Holds the assignment of points. |
envConst_ | is the environment object. |
Definition at line 2263 of file Zoltan2_TaskMapping.hpp.
|
inline |
Constructor. Instead of Solution we have two parameters, numparts.
When this constructor is called, in order to calculate the communication metric, the task adjacency graph is created based on the coordinate model input and partitioning of it. If the communication graph is already calculated, use the other constructors.
comm_ | is the communication object. |
machine_ | is the machineRepresentation object. Stores the coordinates of machines. |
model_ | is the input adapter. |
soln_ | is the solution object. Holds the assignment of points. |
envConst_ | is the environment object. |
Definition at line 2550 of file Zoltan2_TaskMapping.hpp.
|
inline |
Constructor The mapping constructor which will also perform the mapping operation. The result mapping can be obtained by –getAssignedProcForTask function: which returns the assigned processor id for the given task –getPartsForProc: which returns the assigned tasks with the number of tasks.
-task_comm_xadj, task_comm_adj, task_communication_edge_weight_ can be provided NULL. In this case all processors will calculate the same mapping. -If task_comm_xadj, task_comm_adj and provided, algorithm will perform rotations and processors will calculate different mappings, and best one will be reduced. -If task_communication_edge_weight_ is provided with task_comm_xadj, task_comm_adj this will be used when cost is calculated. -recursion_depth is a mandatory argument. In the case part_no_array is not null, this parameter should represent the length of part_no_array. If part_no_array is given as NULL, then this will give the recursion depth for the algorith, Maximum number is ceil(log_2(min(num_processors, num_tasks))), and providing a higher number will be equivalant to this. Partitioning algorithm will work as RCB when maximum number is given, which performs the best mapping results. -part_no_array: The best results are obtained when this parameter is given as NULL. But if this is provided, partitioning will use this array for partitioning each dimension to the given numbers. The multiplication of these numbers should be equal to min(num_processors, num_tasks). -machine_dimensions: This can be NULL, but if provided the algorithm will perform shift of the machine coords so that the largest gap is treated as wrap-around link.
env_const_ | the environment object. |
problemComm | is the communication object. |
proc_dim | dimensions of the processor coordinates. |
num_processors | is the number of processors |
machine_coords | is the coordinates of the processors. |
task_dim | is the dimension of the tasks. |
num_tasks | is the number of tasks. |
task_coords | is the coordinates of the tasks. |
task_comm_xadj | is the task communication graphs xadj array. (task i adjacency is between task_comm_xadj[i] and task_comm_xadj[i + 1]) |
task_comm_adj | is task communication graphs adj array. |
task_communication_edge_weight_ | is the weight of the communication in task graph. |
recursion_depth | is the recursion depth that will be applied to partitioning. If part_no_array is provided, then it is the length of this array. |
part_no_array | if part_no_array is provided, partitioning algorithm will be forced to use this array for partitioning. However, the multiplication of each entries in this array should be equal to min(num_processors, num_tasks). |
*machine_dimensions,: | the dimensions of the machine network. For example for hopper 17x8x24 This can be NULL, but if provided the algorithm will perform shift of the machine coords so that the largest gap is treated as wrap-around link. |
Definition at line 2890 of file Zoltan2_TaskMapping.hpp.
|
inlineprotected |
doMapping function, calls getMapping function of communicationModel object.
Definition at line 1810 of file Zoltan2_TaskMapping.hpp.
|
inlineprotected |
creates and returns the subcommunicator for the processor group.
Definition at line 1837 of file Zoltan2_TaskMapping.hpp.
|
inlineprotected |
finds the lowest cost mapping and broadcasts solution to everyone.
Definition at line 1904 of file Zoltan2_TaskMapping.hpp.
|
inlineprotected |
Definition at line 1947 of file Zoltan2_TaskMapping.hpp.
|
inlineprotected |
Definition at line 2016 of file Zoltan2_TaskMapping.hpp.
|
inline |
Definition at line 2206 of file Zoltan2_TaskMapping.hpp.
|
inlinevirtual |
Mapping method.
Reimplemented from Zoltan2::Algorithm< Adapter >.
Definition at line 2212 of file Zoltan2_TaskMapping.hpp.
|
inline |
Definition at line 2235 of file Zoltan2_TaskMapping.hpp.
|
inlinevirtual |
Returns the number of parts to be assigned to this process.
Implements Zoltan2::PartitionMapping< Adapter >.
Definition at line 3018 of file Zoltan2_TaskMapping.hpp.
|
inline |
Using the machine dimensions provided, create virtual machine coordinates by assigning the largest gap to be as the wrap around link.
machine_dim,: | the number of dimensions in the machine network. |
machine_dimensions,: | the dimension of the machien network. For example for hopper, 17,8,24 |
numProcs,: | the number of allocated processors. |
mCoords,: | allocated machine coordinates. |
Definition at line 3032 of file Zoltan2_TaskMapping.hpp.
|
inlinevirtual |
getAssignedProcForTask function, returns the assigned tasks with the number of tasks.
procId | procId being queried. |
numProcs | (output), the number of processor the part is assigned to. |
procs | (output), the list of processors assigned to given part.. |
Implements Zoltan2::PartitionMapping< Adapter >.
Definition at line 3138 of file Zoltan2_TaskMapping.hpp.
|
inline |
getAssignedProcForTask function, returns the assigned processor id for the given task
taskId | taskId being queried. |
Definition at line 3148 of file Zoltan2_TaskMapping.hpp.
|
inlinevirtual |
getAssignedProcForTask function, returns the assigned tasks with the number of tasks.
procId | procId being queried. |
numParts | (output), the number of parts the processor is assigned to. |
parts | (output), the list of parts assigned to given processor.. |
Implements Zoltan2::PartitionMapping< Adapter >.
Definition at line 3159 of file Zoltan2_TaskMapping.hpp.
|
inline |
Definition at line 3169 of file Zoltan2_TaskMapping.hpp.
|
protected |
Definition at line 1783 of file Zoltan2_TaskMapping.hpp.
|
protected |
Definition at line 1787 of file Zoltan2_TaskMapping.hpp.
|
protected |
Definition at line 1791 of file Zoltan2_TaskMapping.hpp.
|
protected |
Definition at line 1795 of file Zoltan2_TaskMapping.hpp.
|
protected |
Definition at line 1798 of file Zoltan2_TaskMapping.hpp.
|
protected |
Definition at line 1799 of file Zoltan2_TaskMapping.hpp.
|
protected |
Definition at line 1800 of file Zoltan2_TaskMapping.hpp.
|
protected |
Definition at line 1801 of file Zoltan2_TaskMapping.hpp.
|
protected |
Definition at line 1802 of file Zoltan2_TaskMapping.hpp.
|
protected |
Definition at line 1803 of file Zoltan2_TaskMapping.hpp.
|
protected |
Definition at line 1804 of file Zoltan2_TaskMapping.hpp.