Version 12 (modified by jhammond, 9 years ago) (diff) |
---|
# Summary
OSPRI is a new one-sided runtime communication system for leadership-class supercomputers. It aims to provide more flexible semantics compared to existing one-sided communication systems (e.g. ARMCI) and target modern architectural trends by providing a thread-safe implementation.
Unlike its predecessor, A1, OSPRI will provide only functionality above and beyond what is available MPI, hence is designed to run side-by-side with MPI, not as an independent runtime.
The proper citation of OSPRI is:
OSPRI: J. R. Hammond, J. Dinan, P. Balaji, I. Kabadshow, S. Potluri, and V. Tipparaju, The 6th Conference on Partitioned Global Address Space Programming Models (PGAS). Santa Barbara, CA, October 2012. "OSPRI: An Optimized One-Sided Communication Runtime for Leadership-Class Machines."
The corresponding Bibtex is:
@inproceedings{Hammond:PGAS:2012:OSPRI, author = {Jeff R. Hammond and James Dinan and Pavan Balaji and Ivo Kabadshow and Sreeram Potluri and Vinod Tipparaju}, title = {{OSPRI}: An Optimized One-Sided Communication Runtime for Leadership-Class Machines}, booktitle = {The 6th Conference on Partitioned Global Address Space Programming Models ({PGAS})}, month = oct, year = {2012}, location = {Santa Barbara, CA, USA}, url = {http://www.mcs.anl.gov/publications/paper_detail.php?id=1545}, }
Please see the [preprint](http://www.mcs.anl.gov/publication/ospri-optimized-one-sided-communication-runtime-leadership-class-machines) for details.
# Goals of the project
- maximum communication performance
- scalability to millions of processes
- thread-safety
- support for hybrid systems (e.g. GPUs)
- modular design ala MPI
- software quality (readable, documented source code and robust build system)
- productive interaction with vendors
# Developers
Jeff Hammond, Sayan Ghosh, Sreeram Potluri, and Pavan Balaji have all contributed to this project.