Group Topologies

So far we have treated groups as abstract entities. However, in TACO all object groups are based on dynamic distributed data structures and collective operations are following a parallel visitor pattern. All collective methods assume that the members of a group provide an STL style iterator that (recursively) iterates all members of the group. These iterators are provided by topology classes, which can be explicitly specified:

class FlockMember: public Sheep,
    public GroupMember<FlockMember,BalancedBinTree> {
public:
    FlockMember(int rank) : Sheep (...) { }
};

The declaration above describes the group topology to be a balanced binary tree. Figure [*] shows the execution of reduce() operation on a binary tree. The numbers at the edges indicate logical time steps.

  

Figure: Reduction over a binary tree

Each functor is first propagated asynchronously to the other members of the group in the order defined by the iterator before it is applied to the object instance currently visited. In case of reductions we use the asynchronous apply() method in conjunction with lazy synchronization to synchronize on the availability of results. All results are collected in exactly the same order in which the functor has been spread. Thus, the behaviour remains deterministic and application programmers can get control over collective operations either through the iterator method of the topology class or by means of custom iterators defined by application classes.