We say that a computational artifact is “portable” if it can be implemented on a variety of targets. In this paper, the portable artifact is a strategy for the placement of data on a graphics processing unit (GPU). The strategy is lazy in that it defers at least some placement decisions until the program has accumulated performance data from the current run.

The authors’ system, PORPLE, has three major components: MSL (a small language in which to describe GPU memory architecture), PORPLE-C (a compiler that understands data access patterns), and PLACER (an engine that makes the placement decisions). All of these components run on the central processing unit (CPU) rather than the GPU.

PORPLE-C analyzes the user’s program, creating a description of the data access patterns and a target program that is able to accept data placement suggestions from PLACER and run correctly no matter what those suggestions are. PLACER makes its suggestions based on an encoding of the MSL specification for the GPU, the data access patterns produced by PORPLE-C, and information extracted as the program runs.

PORPLE was evaluated on three different machines, with different GPUs, by comparing it with a rule-based approach. It was consistently able to find GPU data placements that yielded excellent performance of the running program.

The paper is carefully written, gives good descriptions of the PORPLE components, and provides a convincing evaluation. There are extensive references. In order to fully appreciate the details, however, the reader needs to have an understanding of GPU architecture.