Loop Parallelization using Cloud MapReduce for Scientific Workloads

• Grant number: 18/21695-0
• Support Opportunities: Scholarships in Brazil - Post-Doctoral
• Start date: December 01, 2018
• End date: October 31, 2019
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Guido Costa Souza de Araújo
• Grantee: Hervé Yviquel
• Host Institution: Instituto de Computação (IC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil

Abstract

The goal of this project is the creation of a compiler-generated parallel code runtime model for remote, distributed execution in computer clusters of the cloud. The ideas here discussed aim at implementing and consolidating OpenMR, a novel programming model, proposed at IC-Unicamp, that extends the existing Open Multi Processing (OpenMP) Application Program Interface (API), to allow a programmer to transparently parallelise loops for execution in the cloud using MapReduce. The preliminary implementation of an OpenMR prototype demonstrated that the proposal is feasible, as it produced encouraging experimental results, and pointed to a number of novel opportunities for future work, which will de developed under this project. One of the attractiveness of the OpenMR model is the simplicity of its use. Its goal is to provide a seamless programming model to the application programmer. An important insight from prototyping OpenMR is that the cloud can be modeled as an additional level in the computing hierarchy. The idea of a computing hierarchy has been established over many years through the development of computer architecture and a software stack that together execute the code of a computer program. Nevertheless, distributed computation in the cloud has not yet been integrated into this stack because there is a significant utilization gap that hinders the transparent execution of code remotely. For instance, an open problem that must be addressed to better integrate the cloud into this hierarchy is the data movement. With higher communication costs and potentially different programming models, algorithms to optimize this data movement must be investigated. Another problem is related to the investigation of optimization techniques which can hide communication latency by means of pipelined computation, and other parallelisation algorithms.