62. MATE - AN IMPROVED BAMBOO TRANSLATOR FOR LARGE-SCALE LATENCY-TOLERANT MPI APPLICATIONS
Name: Sergio Martin
Grad Year: 2019
We present MATE, a source-to-source C/C++ translator for restructuring MPI applications to hide communication delays, which are a significant bottleneck in large scale computing. The motivation for MATE is to avoid static relocation and inlining of source code, a technique used in a previous translator, called Bamboo. Although Bamboo has been shown to improve performance in large-scale applications, its translation rationale penalizes translation time and also obfuscates code debugging. With MATE, we overcome these limitations by injecting a minimal set of routines that schedules execution of code regions dynamically on runtime. This improvement enables the translation of recursive code, simplifies debugging, and reduces compilation time and code size bloating.