System for allocating messages between virtual channels to avoid deadlock and to optimize the amount of message traffic on each type of virtual channel

Claim 3

The method of claim 1 further comprising the step of placing the defined datelines to decrease imbalances in the utilization of the two types of virtual channels for each processing element node and to avoid deadlock.

Claim 4

The method of claim 1 wherein the messages communicated between the processing element nodes over the physical communication paths include request and response information, and wherein the step of defining two types of virtual channels includes assigning two types of virtual channel buffers capable of storing request information and assigning two types of virtual channel buffers capable of storing response information.

Claim 5

The method of claim 1 further comprising the steps of partitioning the processing element nodes into at least two portions for each of the n dimensions and placing the defined datelines at boundaries of the at least two portions.

Claim 6

The method of claim 4 wherein the allocating steps are performed so that messages between the two types of virtual channels within the at least two portions are allocated prior to allocating messages between the two types of virtual channels crossing portion boundaries.

Claim 7

The method of claim 5 wherein allocating messages between the two types of virtual channels within the at least two portions includes alternating the allocation between the two types of virtual channels depending on the lengths of the associated physical communication paths.

Claim 8

The method of claim 5 wherein allocating messages between the two types of virtual channels crossing portion boundaries includes allocating all virtual channel buffers assigned to physical communication paths which pass one or more processing element node after crossing portion boundaries on a first side to the first type of virtual channel and allocating all other virtual channel buffers assigned to physical communication paths crossing portion boundaries on the first side to the second type of virtual channel, and allocating all virtual channel buffers assigned to physical communication paths which pass one or more processing element node after crossing portion boundaries on a second side to the second type of virtual channel and allocating all other virtual channel buffers assigned to physical communication paths crossing portion boundaries on the second side to the first type of virtual channel.

Claim 9

A multiprocessor computer system comprising: a plurality of processing element nodes; an interconnection network interconnecting the plurality of processing element nodes in an n-dimensional topology, the network including physical communication paths connecting each of the plurality of processing element nodes to adjacent processing element nodes, and two types of virtual channels for each of the n dimensions, each type of said virtual channel having virtual channel buffers assigned to each physical communication path, the virtual channel buffers being capable of storing messages communicated between the processing element nodes over the physical communication paths; and a look-up table storing information indicative of a defined dateline in each one of said virtual channels, said dateline representing a communication link between two virtual channel buffers, which complete a cyclic path in their associated virtual channel, wherein if messages cross the dateline a cyclic buffer dependency can occur which creates a deadlock, the look-up table further storing information indicative of an allocation of messages to the virtual channel buffers in any one of the two types of virtual channels when the messages are to be transferred among the processing element nodes in any one of the n dimensions without crossing a dateline, and an allocation of messages such that messages cannot cross a dateline in its associated type of virtual channel, but instead must use the other type of virtual channel to cross the dateline to thereby avoid deadlock.

Claim 10

A multiprocessor computer system comprising: a plurality of processing element nodes; an interconnection network interconnecting the plurality of processing element nodes in an n-dimensional topology, the network including physical communication paths connecting each of the plurality of processing element nodes to adjacent processing element nodes, and two types of virtual channels for each of the n dimensions, each type of said virtual channel having virtual channel buffers assigned to each physical communication path, the virtual channel buffers being capable of storing messages communicated between the processing element nodes over the physical communication paths; and a look-up table storing information indicative of a defined dateline in each one of said virtual channels, said dateline representing a communication link between two virtual channel buffers, which complete a cyclic path in their associated virtual channel, wherein if messages cross the dateline a cyclic buffer dependency can occur which creates a deadlock, the look-up table further storing information indicative of an allocation of messages to the virtual channel buffers in any one of the two types of virtual channels when the messages are to be transferred among the processing element nodes in any one of the n dimensions without crossing a dateline, and an allocation of messages such that messages cannot cross a dateline in its associated type of virtual channel, but instead must use the other type of virtual channel to cross the dateline to thereby avoid deadlock.

Claim 11

The method of claim 1 wherein the allocating steps are performed to decrease imbalances in the utilization of the two types of virtual channels for each processing element node.