A Transformation to Convert Packing Code to Compact Datatypes for Efficient Zero-Copy Data Transfer

Many high performance applications spend considerable timepacking data into contiguous communication buffers. Datatypesprovide an alternative by describing the layout of the communi-cated data. This empowers the runtime system to retrieve non-contiguous elements directly from application data structures.However, programmers find complex datatypes hard to use andare reluctant to invest time and effort to rewrite packing code todatatype code. Fortunately, the transformation from packing codeto datatypes can be automated, and the programmer can replacepacking code with datatypes at the push of a button. The transfor-mation allows easy porting of applications to new machines thatbenefit from datatypes, thus improving programmer productivity.We present an algorithm for converting packing code to datatypecode based on a novel IR and a suite of optimizations. We have im-plemented the algorithm in a tool that transforms C packing codeto an MPI datatype, and rewrites the packing code consumer toinstead use the datatype. Our evaluation shows that our algorithmis applicable to real-world packing code, that it is fast enough to beused interactively, and that the datatypes it produces are compactand well optimized. Finally, we evaluate the performance of thecode produced by our tool, showing that it outperforms the orig-inal packing code on a state-of-the art system. However, the realbenefit of datatypes is in the future, when hardware support fornon-contiguous gather-scatter transfers becomes ubiquitous.