Multidestination message passing in wormhole k-ary n-cube networks with base routing conformed paths

Abstract

This paper proposes multidestination message passing on wormhole k-ary n-cube networks using a new base-routing-conformed-path (BRCP) model. This model allows both unicast (single-destination) and multidestination messages to co-exist in a given network without leading to deadlock. The model is illustrated with several common routing schemes (deterministic, as well as adaptive), and the associated deadlock-freedom properties are analyzed. Using this model a set of new algorithms for popular collective communication operations, broadcast and multicast, are proposed and evaluated. It is shown that the proposed algorithms can considerably reduce the latency of these operations compared to the Umesh (unicast-based multicast) 22] and the Hamiltonian-Path-based 21] schemes. A very interesting result that is presented shows that a multicast can be implemented with reduced or near-constant latency as the number of processors participating in the multicast increases beyond a certain number. It is also shown that the BRCP model can take advantage of adaptivity in routing schemes to further reduce the latency of these operations. The multidestination mechanism and the BRCP model establish a new foundation to provide fast and scalable collective communication support on wormhole-routed systems.

Quarc: A Novel Network-On-Chip Architecture

“…After path selection, the intermediate destinations perform absorb-and-forward operations along the path. Hamilton path-based algorithm [5] and the Base Routing Conformed Path (BRCP) approach [1] are examples of path-based algorithms utilizing absorb-and-forward property at hardware layer.…”

Trade-Offs between Latency, Complexity, and Load Balancing with Multicast Algorithms

“…Ensuring a scalable implementation of a wide variety of parallel applications necessitates efficient implementation of multicast communication. In general, the literature outlines three main approaches to deal with the multicast problem: unicast-based [1], [3], tree-based [3], [17], and path-based [2], [3], [8], [14], [15], [18]. A number of studies have shown that path-based algorithms exhibit superior performance characteristics over their unicast and tree-based counterparts [2], [14], [15], [18].…”

“…In path-based multicast, when the flits of a message reach one of the destination nodes in the multicast group, they are copied to local memory while they continue to flow through the node to reach the other destinations [2], [3], [8], [18]. The message is removed from the network when it reaches the last destination in the multicast group.…”

“…Multicast, in which a source node sends the same message to an arbitrary number of destination nodes in the network, is one of the most useful collective communication operations [1], [2], [3], [6]. Due to its extensive use, efficient multicast is critical to the overall performance of parallel machines [1], [2], [3], [4], [14], [18]. For instance, multicast is frequently used by many important applications such as parallel search and parallel graph algorithms [3], [14].…”

“…Nevertheless, n-dimensional meshes have undoubtedly been the most popular interconnection network used in practice [2], [3], [5], [6], [9], [10], [11], [20], [21] due to their desirable topological properties including ease of implementation, modularity, low diameter, and ability to exploit locality exhibited by many parallel applications [3]. Meshes are suited to a variety of applications including matrix computation, image processing, and problems whose task graphs can be embedded naturally into the topology [3], [6], [10], [18], [19]. Meshes have been used in a number of real parallel machines including the Intel Paragon, MIT J-machine, Cray T3D, T3E, Caltech Mosaic, Intel Touchstone Delta, and Stanford DASH [3].…”