Multiprocessor computer system
AssigneeNCR Corporation
InventorsPhilip M. Neches, Robert J. McMillen, M. Cameron Watson, David J. Chura
Life of this patent
Estimating the term…
Abstract
A multistage interconnect network (MIN) capable of supporting massive parallel processing, including point-to-point and multicast communications between processor modules (PMs) which are connected to the input and output ports of the network. The network is built using interconnected switch nodes arranged in 2 log.sub.b N stages, wherein b is the number of switch node input/output ports, N is the number of network input/output ports and log.sub.b N indicates a ceiling function providing the smallest integer not less than log.sub.b N. The additional stages provide additional paths between network input ports and network output ports, thereby enhancing fault tolerance and lessening contention.
Claims6
A multiprocessor system, comprising:
- (a) a plurality of processors;
- (b) a network for connecting the processors, wherein the network comprises a plurality of switch nodes arranged into more than log.sub.b N switch node stages, wherein b is a total number of switch node input/output ports, N is a total number of network input/output ports, and log.sub.b N indicates a ceiling function providing the smallest integer not less than log.sub.b N, the switch node stages thereby providing a plurality of paths between any network input port and network output port to enhance fault tolerance and lessen contention, the switch node stages being configured to provide a plurality of bounce-back points at a highest switch node stage of the network for directing transmissions of messages through the network, the bounce-back points logically differentiating between switch nodes that load balance messages through the network from switch nodes that direct messages to receiving processors;
- (c) means for partitioning the processors into one or more superclusters comprising logically independent subsets of processors;
- and
- (d) means for multicast communicating within superclusters by steering a multicast message to a particular bounce-back point in the network for subsequent routing to the processors in the supercluster, thereby preventing deadlock in the network by permitting only one multicast message through the particular bounce-back point at a time and preventing multicast messages to different superclusters from interfering with each other.
Dependencies are inferred from the claim text.