Method and system for improving computational concurrency using a multi-threaded GPU calculation engine

摘要

A computing system may include a plurality of GPUs, one or more CPUs, and a data store containing data to be analyzed. A user may register a calculation to be performed on the data, and the system may determine a scheme for allocating portions of the calculation and subsets of the data to different GPUs for concurrent execution. The system may also modify a user-provided algorithm to convert portions of the algorithm that are path-dependent into atomic path-independent operations that may be performed by GPU threads asynchronously. The system may also determine coordination operations sufficient for merging sub-results generated from individual GPU threads' execution of the atomic operations back into a path-dependent computation. The system may be configured to use any number of similar or different GPU devices, and for GPU devices to be added to or removed from the system without requiring programmers to alter their algorithms.