Abstract: A method includes obtaining a compute graph for computing a first tensor, identifying in the graph a reduction operation in at least one dimension of the first tensor, locating, at the operation, a cut point that cuts the graph into first and second portions, and determining a plurality of slices of the first tensor. The method also includes backpropagating the cut point through the graph to define a plurality of first graph pieces for the first portion, each particular first graph piece representing a computation of a particular slice of the plurality of slices based on a particular portion of a plurality of portions of a second tensor. The method further includes defining one or more second graph pieces to combine outputs of the first graph pieces, and executing the first graph pieces and the second graph pieces to execute the first portion of the compute graph.

Abstract: Aspects of the disclosure are directed to protecting non-public data used in computer-based trials. Reports can be generated based on the computer-based trials that contain performance metrics or other data points to evaluate the computer-based trials. The reports can be viewed by data providers and/or trial providers without divulging the non-public data. Data providers can provide the non-public data for running the computer-based trials. Trial providers can run the computer-based trials. A cloud provider can provide infrastructure for storing source code for the computer-based trials, the non-public data, and reports generated from the computer-based trials. The cloud provider can also provide infrastructure for executing the computer-based trials and generating the reports from the computer-based trials.

Abstract: A method includes obtaining a graph that includes a plurality of nodes each representing a computational job and connected to one or more other nodes via edges. Each edge represents a dependency between two connected nodes. For each computing device of a distributed computing system, the method includes obtaining resource characteristics. The method includes slicing, based on the dependencies of connected nodes, the graph into a plurality of sub-graphs each including one or more nodes. For each respective sub-graph, the method includes determining a computational cost of the respective sub-graph and distributing, based on the computational cost and the resource characteristics, the respective sub-graph to the respective one of the computing devices. The method includes receiving a respective result from the respective one of the computing devices based on execution of the computational jobs. The method includes determining a total result based on each respective result.