US 2016/0105262 A1 · COLLECTIVE NETWORK FOR COMPUTER STRUCTURES

Abstract

A system and method for enabling high-speed, low-latency global collective communications among interconnected processing nodes. The global collective network optimally enables collective reduction operations to be performed during parallel algorithm operations executing in a computer structure having a plurality of the interconnected processing nodes. Router devices are included that interconnect the nodes of the network via links to facilitate performance of low-latency global processing operations at nodes of the virtual network and class structures. The global collective network may be configured to provide global barrier and interrupt functionality in asynchronous or synchronized manner. When implemented in a massively-parallel supercomputing structure, the global collective network is physically and logically partitionable according to needs of a processing algorithm.

Claims20

Dependent claims

A method of correcting reduction operations in a presence of packet corruption in a multi-node computer network, the method comprising:

a first node of the network sending an operand packet to a second node of the network over a link between the first and second nodes, said packet including one or more operands;
the first node keeping a copy of the packet on a sending end of the link until the first node receives an acknowledgment from the second node that the packet was received by the second node without error;
the second node testing the packet to determine if the packet is error free;
and
if the second node determines that the packet is not error free, then the second node setting a flag in the packet to mark the packet as corrupt, returning an acknowledgement to the first node which specifies that the packet was received with error, and requesting retransmission of the packet;
returning the link to a known state and retransmitting the packet to the second node;
and
when the re-transmitted packet arrives at the second node, repeating a reduction computation utilizing stored copies of uncorrupted packets.

5 more elements

A computer readable program storage device readable by a computer, tangibly embodying a program of instructions executable by the computer to perform method steps for correcting reduction operations in a presence of packet corruption in a multi-node computer network, the method steps comprising:

a first node of the network sending an operand packet to a second node of the network over a link between the first and second nodes, said packet including one or more operands;
the first node keeping a copy of the packet on a sending end of the link until the first node receives an acknowledgment from the second node that the packet was received by the second node without error;
the second node testing the packet to determine if the packet is error free;
and
if the second node determines that the packet is not error free, then the second node setting a flag in the packet to mark the packet as corrupt, returning an acknowledgement to the first node which specifies that the packet was received with error, and requesting retransmission of the packet;
when the re-transmitted packet arrives at the second node, the second node repeating a reduction computation utilizing stored copies of uncorrupted packets.

3 more elements

Dependencies are inferred from the claim text.

Collective network for computer structures

Life of this patent

Collective network for computer structures

Life of this patent

Abstract

Claims20