Pipelining protocols in misaligned buffer cases
AssigneeInternational Business Machines Corporation
InventorsCharles J. Archer, Michael Blocksome, Bob R. Cernohous, Joseph D. Ratterman, Brian E. Smith
Life of this patent
Estimating the term…
Abstract
Systems, methods and articles of manufacture are disclosed for effecting a desired collective operation on a parallel computing system that includes multiple compute nodes. The compute nodes may pipeline multiple collective operations to effect the desired collective operation. To select protocols suitable for the multiple collective operations, the compute nodes may also perform additional collective operations. The compute nodes may pipeline the multiple collective operations and/or the additional collective operations to effect the desired collective operation more efficiently.
Claims12
A non-transitory computer-readable storage medium containing a program which, when executed, performs a specified collective operation on a parallel computer comprising a plurality of compute nodes, the specified collective operation comprising:
- performing a first collective operation by a source node on a plurality of target nodes to determine a characteristic of the target nodes, wherein the characteristic for each target node specifies whether a buffer of the respective target node is N-byte aligned, where N is a power of two;
- from a plurality of protocols, selecting a target protocol based on determining whether all of the target nodes are N-byte aligned;
- performing the specified collective operation by the source node on the plurality of target nodes using the target protocol;
- performing a second collective operation by one of the target nodes on a plurality of downstream nodes to determine a characteristic of each of the downstream nodes, wherein the characteristic for each downstream node specifies whether a buffer of the respective downstream node is N-byte aligned, where N is a power of two;
- from the plurality of protocols, selecting a downstream protocol based on determining whether all of the downstream nodes are N-byte aligned;
- and
- performing the specified collective operation by the one of the target nodes on the plurality of downstream nodes using the downstream protocol.
A parallel computing system, comprising:
- a source node, having at least a processor and a memory;
- and
- a plurality of target nodes, each having at least a processor and a memory, wherein the nodes are configured to perform an operation, comprising:
- performing a first collective operation by the source node on the plurality of target nodes to determine a characteristic of the target nodes, wherein the characteristic for each target node specifies whether a buffer of the respective target node is N-byte aligned, where N is a power of two,
- from a plurality of protocols, selecting a target protocol based on determining whether all of the target nodes are N-byte aligned,
- performing a specified collective operation by the source node on the plurality of target nodes using the target protocol,
- performing a second collective operation by one of the target nodes on a plurality of downstream nodes to determine a characteristic of each of the downstream nodes, wherein the characteristic for each downstream node specifies whether a buffer of the respective downstream node is N-byte aligned, where N is a power of two,
- from the plurality of protocols, selecting a downstream protocol based on determining whether all of the downstream nodes are N-byte aligned, and
- performing the specified collective operation by the one of the target nodes on the plurality of downstream nodes using the downstream protocol.
Dependencies are inferred from the claim text.