Embedding global and collective in a torus network with message class map based tree path selection
Embodiments of the invention provide a method, system and computer program product for embedding a global barrier and global interrupt network in a parallel computer system organized as a torus network. The computer system includes a multitude of nodes. In one embodiment, the method comprises taking inputs from a set of receivers of the nodes, dividing the inputs from the receivers into a plurality of classes, combining the inputs of each of the classes to obtain a result, and sending said result to a set of senders of the nodes. Embodiments of the invention provide a method, system and computer program product for embedding a collective network in a parallel computer system organized as a torus network. In one embodiment, the method comprises adding to a torus network a central collective logic to route messages among at least a group of nodes in a tree structure.
1. A method of embedding a collective network in a parallel computer system organized as a torus network, said computer system including a multitude of nodes, each of the nodes having a plurality of receivers and a plurality of senders, the method comprising:
adding to the torus network a central collective logic to route messages among at least a group of said nodes in a tree structure, wherein, at defined times, one of said group of nodes is a root node and the others of said group of nodes are leaf or intermediate nodes, and including
said central collective logic forming separate up-tree and down-tree paths for routing messages among said group of nodes;
the central collective logic selecting one of the up-tree path and down-tree path for one of the messages for each sender depending on a collective class map of said one of the messages;
routing messages from the leaf or intermediate nodes to the root node in an up tree direction; processing the messages being routed from the leaf or intermediate nodes to the root node to form a processed message; and
sending the processed message back from the root node to at least one of the leaf, intermediate or root nodes.
2. The method according to claim 1 , wherein said processing includes combining the messages from the leaf or intermediate nodes into one, combined message.
3. The method according to claim 1 , wherein the sending includes sending the processed message from the root node to all the leaf, intermediate and root nodes in a down tree direction.
4. The method according to claim 1 , wherein said adding further includes:
adding dedicated injection fifos and dedicated reception fifos for collective operations; and
adding dedicated collective virtual channels to each of the receivers, and adding a central collective logic block to handle arbitration and collective data flow.
5. The method according to claim 1 , wherein one can double the number of classes when the size of a torus dimension is two.
6. A system for embedding a collective network in a parallel computer system organized as a torus network, said computer system including a multitude of nodes, each of the nodes having a plurality of receivers and a plurality of senders, the system comprising one or more processing units configured for:
providing a central collective logic to the torus network to route messages among at least a group of said nodes in a tree structure, wherein, at defined times, one of said group of nodes is a root node and the others of said group of nodes are leaf or intermediate nodes, and including
said central collective logic forming separate up-tree and down-tree paths for routing messages among said group of nodes;
the central collective logic selecting one of the up-tree path and down-tree path for one of the messages for each sender depending on a collective class map of said one of the messages;
routing messages from the leaf or intermediate nodes to the root node in an up tree direction;
processing the messages being routed from the leaf or intermediate nodes to the root node to form a processed message; and
sending the processed message back from the root node to at least one of the leaf, intermediate or root nodes.
7. The system according to claim 6 , wherein said processing includes combining the messages from the leaf, intermediate and root nodes into one, combined message.
8. The system according to claim 6 , wherein the sending includes sending the processed message from the root node to all the leaf, intermediate and root nodes in a down tree direction.
9. The system according to claim 8 , wherein said adding further includes adding dedicated injection fifos and dedicated reception fifos for collective operations.
10. The system according to claim 9 , wherein said adding further includes adding dedicated collective virtual channels to each of the receivers, and adding a central collective logic block to handle arbitration and collective data flow.