System and method for optimizing data-transfer among multiple compute units in a data-parallel computing system
AssigneeSambaNova Systems, Inc.
InventorsGreg Dykema, Aarti Lalwani
Life of this patent
Estimating the term…
Abstract
System and method for optimizing data-transfer among multiple compute units in a data-parallel computing system. A topological communications configurator (TCC) determines a connections-optimized configuration of processors associated with compute nodes of the computing system. The processors can execute dataflow workers of an application and form intranodal segments of an internodal interconnection topology coupling the intranodal segments. The TCC determines the connections-optimized configuration based on internodal communications costs corresponding to communications routes among the internodal segments via the internodal interconnection fabric.
Claims19
A data processing system, comprising:
- a first computing system comprising a first processor and a topological communications configurator (TCC) comprising a computing program configured to execute on the first processor;
- a second computing system comprising a plurality of compute nodes, compute nodes among the plurality of compute nodes comprising at least one dataflow processor, each of the at least one dataflow processor configured to execute a compute worker among a plurality of compute workers, the plurality of compute workers configured to execute a computing application of the second computing system;
- and
- an internodal fabric configured to communicatively couple compute nodes among the plurality of compute nodes,
- wherein the TCC is configured to:
- determine an internodal interconnection topology of the plurality of compute nodes, the internodal interconnection topology comprising the internodal fabric;
- determine, based on the internodal interconnection topology, internodal communications routes communicatively interconnecting, via the internodal fabric, a set of intranodal segments among a plurality of intranodal segments, each of the plurality of intra nodal segments comprising an intranodal interconnection of dataflow processors included in respective nodes among the plurality of compute nodes, each of the plurality of intranodal segments comprising a set of compute workers, among the plurality of compute workers, executed by the dataflow processors included in a node among the respective nodes, the set of compute workers included in a worker logical topology;
- determine internodal communications costs corresponding to communications routes among the internodal communications routes;
- and
- determine, based on the internodal communications costs, a connections-optimized configuration of interconnected segments among the set of intranodal segments.
A computer-implemented method comprising:
- determining, by a topological communications configurator (TCC) of a first computing system, an internodal interconnection topology of a plurality of compute nodes of a second computing system, the internodal interconnection topology comprising an internodal fabric;
- determining, by the TCC, based on the internodal interconnection topology, internodal communications routes interconnecting, via the internodal fabric, a set of intranodal segments among a plurality of intranodal segments, each of the plurality of intranodal segments comprising an intranodal interconnection of dataflow processors included in respective nodes among the plurality of compute nodes, each of the plurality of intranodal segments corresponding to a respective portion of a worker logical topology comprising compute workers configured to execute an application of the second computing system;
- determining, by the TCC, internodal communications costs corresponding to communications routes among the internodal communications routes;
- and
- determining a connections-optimized configuration of interconnected segments.
A computer program product comprising a computer readable storage medium having first program instructions embodied therewith, wherein the first program instructions are executable by at least one processor to cause the at least one processor to:
- determine an internodal interconnection topology of a plurality of compute nodes of a computing system, the internodal interconnection topology comprising an internodal fabric;
- determine, based on the internodal interconnection topology, internodal communications routes communicatively interconnecting, via the internodal fabric, a set of intranodal segments among a plurality of intranodal segments, each of the plurality of intranodal segments comprising an intranodal interconnection of processors included in respective nodes among the plurality of compute nodes, each of the plurality of intranodal segments corresponding to a respective portion of a worker logical topology comprising compute workers configured to execute an application of the computing system;
- determine internodal communications costs corresponding to communications routes among the internodal communications routes;
- and
- determine, based on the internodal communications costs, a connections-optimized configuration of interconnected segments.
Dependencies are inferred from the claim text.