xgboost.tracker.RabitTracker Class Reference

Public Member Functions
None	__init__ (self, str host_ip, int n_workers, int port=0, bool use_logger=False, str sortby="host")
None	__del__ (self)
Dict[str, Union[str, int]]	worker_envs (self)
List[int]	find_share_ring (self, _TreeMap tree_map, Dict[int, int] parent_map, int rank)
_RingMap	get_ring (self, _TreeMap tree_map, Dict[int, int] parent_map)
Tuple[_TreeMap, Dict[int, int], _RingMap]	get_link_map (self, int n_workers)
None	accept_workers (self, int n_workers)
None	start (self, int n_workers)
None	join (self)
bool	alive (self)

Data Fields
	port
	sock
	host_ip
	n_workers
	thread

Protected Member Functions
Tuple[_TreeMap, Dict[int, int]]	_get_tree (self, int n_workers)
List[WorkerEntry]	_sort_pending (self, List[WorkerEntry] pending)

Static Protected Member Functions
List[int]	_get_neighbor (int rank, int n_workers)

Protected Attributes
	_use_logger
	_sortby

Detailed Description

tracker for rabit

Constructor & Destructor Documentation

__init__()

None xgboost.tracker.RabitTracker.__init__	(		self,
		str	host_ip,
		int	n_workers,
		int	port = 0,
		bool	use_logger = False,
		str	sortby = "host"
	)

A Python implementation of RABIT tracker.

Parameters
..........
use_logger:
    Use logging.info for tracker print command.  When set to False, Python print
    function is used instead.

sortby:
    How to sort the workers for rank assignment. The default is host, but users
    can set the `DMLC_TASK_ID` via RABIT initialization arguments and obtain
    deterministic rank assignment. Available options are:
      - host
      - task

Member Function Documentation

accept_workers()

None xgboost.tracker.RabitTracker.accept_workers	(		self,
		int	n_workers
	)

Wait for all workers to connect to the tracker.

alive()

bool xgboost.tracker.RabitTracker.alive

(

self

)

Wether the tracker thread is alive

find_share_ring()

List[int] xgboost.tracker.RabitTracker.find_share_ring	(		self,
		_TreeMap	tree_map,
		Dict[int, int]	parent_map,
		int	rank
	)

get a ring structure that tends to share nodes with the tree
return a list starting from rank

get_link_map()

Tuple[_TreeMap, Dict[int, int], _RingMap] xgboost.tracker.RabitTracker.get_link_map	(		self,
		int	n_workers
	)

get the link map, this is a bit hacky, call for better algorithm
to place similar nodes together

get_ring()

_RingMap xgboost.tracker.RabitTracker.get_ring	(		self,
		_TreeMap	tree_map,
		Dict[int, int]	parent_map
	)

get a ring connection used to recover local data

join()

None xgboost.tracker.RabitTracker.join

(

self

)

Wait for the tracker to finish.

start()

None xgboost.tracker.RabitTracker.start	(		self,
		int	n_workers
	)

Strat the tracker, it will wait for `n_workers` to connect.

worker_envs()

Dict[str, Union[str, int]] xgboost.tracker.RabitTracker.worker_envs

(

self

)

get environment variables for workers
can be passed in as args or envs

---

The documentation for this class was generated from the following file:

• External/xgboost/python-package/xgboost/tracker.py