OAC Core: Small: Collaborative Research: Scalable Run-Time for Highly Parallel, Heterogeneous Systems

点击翻译按钮获取中文摘要

Supercomputing has become an essential tool in many scientific fields, including advances in engineering and medicine, and contributes to national security. Progress in many areas depends on continued improvements in the performance of supercomputers and their usability. Communication between processes is a critical component of this effort and is the target of this project. This project departs from the traditional communication protocols. Rather, the project focuses on providing middle ground solutions between hardware and software. This approach potentially reduces communication overheads and better matches the functionality of the communication library to the capabilities of modern communication adapters and also improves the match between the requirements of modern parallel computing frameworks and applications. By improving the communication capabilities of computational platforms, this project will promote faster and more flexible communication capabilities and will improve the time to completion of scientific applications. It, therefore, increases the scientific throughput of existing and future cyberinfrastructure platforms. The research and educational outcomes of this project are closely related, resulting in highly trained new generations of researchers and engineers leading to a more efficient and globally competent workforce. Therefore, this project aligns with the NSF's mission to promote the progress of science and to advance national prosperity and welfare through science, and serves the national interest. This project brings together a multidisciplinary team and aims at breaking away from the limitation of standards such as Message Passing Interface and pointing the way for handling the needs of future computational frameworks and high-end systems. To this end the project (1) designs and implements a communication library with new communication primitives to enable fast coordination with no serial bottleneck, to manage irregular, fine grain communication, and to provide new efficient synchronization mechanisms; (2) demonstrates the value of this library by using it to accelerate multiple task-based runtimes (Legion, PaRSEC) and communication libraries (MPI and GasNET); (3) demonstrates the value of hardware support by porting key components to a programmable NIC; and (4) delivers improvements and extensions to mainstream communication libraries to provide the new functionality. This work puts a special emphasis on emerging programming models, such as Legion or PaRSEC, and on emerging application domains, such as graph analytics. It aims at an orthogonal design where different mechanisms for associating producer buffer with consumer buffer can be composed with different mechanisms for synchronizing producer and consumer; and where mechanisms can be specialized so as to allow efficient hardware support.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.