Package org.apache.reef.io.network.group.api.operators

Provides the interfaces for MPI style group communication operations.

See: Description

Interface Summary

Interface	Description
AllGather<T>	MPI AllGather Operator.
AllReduce<T>	MPI All Reduce Operator.
Broadcast	MPI Broadcast operator.
Broadcast.Receiver<T>	Receivers or Non-roots.
Broadcast.Sender<T>	Sender or Root.
Gather	MPI Gather Operator.
Gather.Receiver<T>	Receiver or Root.
Gather.Sender<T>	Senders or non-roots.
GroupCommOperator
Reduce	MPI Reduce operator.
Reduce.Receiver<T>	Receiver or Root.
Reduce.ReduceFunction<T>	Interface for a Reduce Function takes in an Iterable returns an.
Reduce.Sender<T>	Senders or non roots.
ReduceScatter<T>	MPI Reduce Scatter operator.
Scatter	MPI Scatter operator
Scatter.Receiver<T>	Receiver or non-roots.
Scatter.Sender<T>	Sender or Root.

Class Summary

Class	Description
AbstractGroupCommOperator

Package org.apache.reef.io.network.group.api.operators Description

Provides the interfaces for MPI style group communication operations. The interface is asymmetric for asymmetric operations and symmetric for symmetric operations unlike MPI which provides symmetric operations for all operations. The interface is asymmetric in the sense that Senders and Receivers are separated out for operations like Scatter and Gather. All participants do not execute the same function. A sender sends and a receiver receives. The interface only concentrates on the data part because we are on the data-plane of things in REEF. The control information is embedded in the Configuration used to instantiate these operators. It is the responsibility of the Driver, the primary agent in the control-plane to configure these operators, that is, denote who is the sender, who are the receivers, what Codec need to be used and so on for an operation like Scatter with the root node acting as a sender and the other nodes as receivers. One thing implicit in MPI operations is the ordering of processors based on their ranks which determines the order of operations. For ex., if we scatter an array of 10 elements into 10 processors, then which processor gets the 1st entry and so on is based on the rank. In our case we do not have any ranks associated with tasks. Instead, by default we use the lexicographic order of the task ids. These can also be over-ridden in the send/receive/apply function calls