Running MPI Applications over Opportunistic Cloud Infrastructures
Authors:
- Nathalia GarcesNGNathalia Garces
- German A. SoteloGSGerman A. Sotelo
- Mario VillamizarMVMario Villamizar
- Harold CastroHCHarold Castro
- 1 November 2012
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 309-314
Abstract
This paper analyzes in detail the difficulties and issues that must be faced when running MPI (Message Passing Interface) applications over non-dedicated infrastructures. Alternatives to overcome these issues on a specific opportunistic infrastructure, called UnaCloud, are presented. We designed and implemented a UnaCloud extension to allow automatic recovery of the execution of MPI applications. This extension will provide new opportunities for research groups, allowing researchers to run MPI applications with a minimum overhead and at a low cost. We tested our implementation running a large application for several hours: Gromacs MPI, a molecular dynamic application that uses MPI to predict and redefine the tertiary structure of a particular protein. Our results show that running such an application is indeed possible, even if the underlying infrastructure exhibits high volatility. This provides a setting to introduce UnaCloud as an opportunistic platform that can be used to run Bag of Tasks (BoT) and MPI applications.Keywords
This publication has 8 references indexed in Scilit:
- UnaGrid/UnaCloud: A Desktop Grid and Cloud Computing Solution2011
- Bridging the High Performance Computing Gap: the OurGrid Experience2007
- Message passing over windows-based desktop grids2006
- The Lam/Mpi Checkpoint/Restart Framework: System-Initiated CheckpointingThe International Journal of High Performance Computing Applications, 2005
- The implementation of the BSP parallel computing model on the InteGrade Grid middleware2005
- SETI@homeCommunications of the ACM, 2002
- A high-performance, portable implementation of the MPI message passing interface standardParallel Computing, 1996
- A bridging model for parallel computationCommunications of the ACM, 1990