Static Generation of Efficient OpenMP Offload Data Mappings

Increasing heterogeneity in HPC architectures and compiler advancements have led to OpenMP being frequently used to enable computations on heterogeneous devices. However, the efficient movement of data on heterogeneous computing platforms is crucial for achieving high utilization. The implicit OpenMP data-mapping rules often result in redundant data transfer, which can be a bottleneck for program performance. Programmers must explicitly map data between the host and connected accelerator devices to achieve efficient data movement. For this, OpenMP offers the target data and target update constructs. Ensuring efficient data transfer requires programmers to reason about complex data flow. This can be a laborious and error-prone process since the programmer must keep a mental model of data validity and lifetime spanning multiple data environments. Any automated analysis should maximize data reuse, minimize data transfer, and must consider control flow and context from function call sites, making the analysis interprocedural and context sensitive. In this paper, we present a static analysis tool, OMPDart (OpenMP DAta Reduction Tool), for OpenMP programs that models data dependencies between host and device regions and applies source code transformations to achieve efficient data transfer. The analysis is based on a hybrid data structure that joins an Abstract Syntax Tree (AST) with a Control Flow Graph (CFG). Our evaluations on nine HPC benchmarks demonstrate that OMPDart is capable of generating effective data mapping constructs that substantially reduce data transfer between host and device. OMPDart helps reduce data transfers by 85% and improves runtime performance by 1.6x over an expert-defined implementation of LULESH 2.0.