University of Tennessee at Chattanooga
Place of Publication
As supercomputers have approached exascale performance, several scalability issues have emerged within MPI. These issues arise because MPI includes all processes in the World model, which consumes unacceptable amounts of time and resources at large scale. The Sessions model was developed to combat these issues by removing the requirement of MPI_COMM_WORLD, which provides a more scalable method of creating communication groups in large jobs. Additionally, the Sessions model enables the creation of virtual topologies directly from sets of processes allocated to the execution of a parallel application rather than building virtual topologies from an existing communication group such as MPI_COMM_WORLD. For this project, I implemented the Sessions model in ExaMPI, an MPI implementation designed for modularity, extensibility, and understandability. I also created topological variations of several common communication algorithms and topological connection building to further take advantage of the benefits of the Sessions model. I found that using the Sessions model reduces the time and resources used when a large parallel application begins executing. Additionally, I found that using topological connection building and topological communication algorithms is faster than traditional all-to-all connection building in certain situations.
First, I thank my thesis director Dr. Tony Skjellum for his continued mentorship and expertise throughout my time as a researcher at UTC. I also thank Dr. Eleni Panagiotou for providing many of the research skills I needed to complete this thesis. I owe many thanks to Derek Schafer, whose knowledge of C++ and ExaMPI made this thesis possible. Finally, I owe gratitude to Dr. Howard Pritchard for his advice on implementation details and testing. I acknowledge funding received from the Research Experience for Undergraduates (REU) program through supplements to NSF funding at the University of Tennessee at Chattanooga, under grants # 1822191, 1821926, 1821431, and DMS-1913180.
B. S.; An honors thesis submitted to the faculty of the University of Tennessee at Chattanooga in partial fulfillment of the requirements of the degree of Bachelor of Science.
Electronic data processing; Parallel programs (Computer programs)
Computer Sciences | Software Engineering
v, 70 leaves
Herschberg, Tom, "Simplifying the creation of virtual topologies using MPI Sessions" (2021). Honors Theses.