Collectives

Collective operations

The collective component vt::collective::CollectiveAlg, accessed via vt::theCollective() implements active-message-based distributed collectives over the vt runtime. It performs asynchronous reductions, scatters, barriers, and allows one to safely use MPI interspersed through vt code, while running in a handler.

Reductions

vt comes with several reduction operators built in that can operate over types where the operator already applies (like integers, doubles, etc.) or user-defined overloaded operators. These operators also have specializations for std::vector<T> and std::array<T>.

Operator	Reference
operator&&	`vt::collective::AndOp`
operator\|\|	`vt::collective::OrOp`
operator+	`vt::collective::PlusOp`
std::min	`vt::collective::MinOp`
std::max	`vt::collective::MaxOp`
operator&	`vt::collective::BitAndOp`
operator\|	`vt::collective::BitOrOp`
operator^	`vt::collective::BitXorOp`
<no-operator>	`vt::collective::NoneOp`

A Simple Reduction

#include <vt/transport.h>

// Reduce ints
struct ReduceDataMsg : vt::collective::ReduceTMsg<int> {
  explicit ReduceDataMsg(int val)
    : vt::collective::ReduceTMsg<int>(val)
  { }
};

// Handler to target for reduction
struct ReduceResult {
  void operator()(ReduceDataMsg* msg) {
    auto num_nodes = vt::theContext()->getNumNodes();
    auto output = msg->getConstVal();
    fmt::print("reduction value={}\n", output);
    vtAssert(num_nodes * 50 == output, "Must be equal);
  }
};

int main(int argc, char** argv) {
  vt::initialize(argc, argv);

  auto reduce_msg = vt::makeMessage<ReduceDataMsg>(50);

  NodeType const root_reduce_node = 0;
  vt::theCollective()->global()->reduce<vt::collective::PlusOp<int>,ReduceResult>(
    root_reduce_node, reduce_msg.get()
  );

  vt::finalize(); // spins in scheduler until termination
  return 0;
}