High-performance and scalable communication subsystems for exascale computing

高性能计算（HPC）是许多科学发现和工程创新的关键。它被用于解决各种领域的计算密集型问题，如药物发现、全球气候系统建模、石油和天然气的地震处理、绿色能源、基因组学和生物信息学以及天体物理学。科学/工程仿真主要使用消息传递接口（MPI）库编写。这些模拟中的并行进程在通过MPI库进行广泛通信的同时对其本地数据进行计算。通信对运行在HPC集群上的MPI应用程序的性能和可伸缩性有不利影响。随着多核和即将出现的多核架构在所有级别上提供越来越多的并行性，由数十万个节点和数百万个核心组成的高性能计算集群和复杂的网络拓扑在未来几年有望突破每秒10^18次浮点运算的障碍。随着这种高度分层集群的出现，MPI必须针对性能和可扩展性进行优化，以应对不断增长的大规模模拟需求。所提议的研究是高度原创和创新的，因为它解决了MPI中的关键问题，包括过程映射的拓扑感知，为合作伙伴进程合并专用队列，服务质量提供合作伙伴/非合作伙伴流量，以及开发基于光纤的异步进展技术。这项研究的结果将与加拿大的各个部门相关，包括加拿大环境部、加拿大计算机部、加拿大基因组科学中心、石油和天然气行业，并最终与加拿大公众有关。预计本研究结果将对目标群体产生重大影响，并为未来的研究带来新的方向。所提出的研究是培训HQP的理想选择，因为它有一个坚实的基础，可以立即转化为实际应用和实施。HPC和网络领域对毕业生的需求很大，受过HQP培训的学生将在学术界和工业界竞争工作。

High-Performance Computing (HPC) is the key to many scientific discoveries and engineering innovations. It is used to tackle computationally-intensive problems in fields as diverse as drug discovery, modeling of global climate system, seismic processing for oil and gas, green energy, genomics and bioinformatics, and astrophysics. Scientific/engineering simulations are mainly written with the Message-Passing Interface (MPI) library. Parallel processes in these simulations compute on their local data while extensively communicating with each other through the MPI library. Communication adversely affects the performance and scalability of MPI applications running on HPC clusters. With the availability of multi-core and soon many-core architectures offering increasing parallelism at all levels, HPC clusters consisting of hundreds of thousands of nodes with millions of cores and complex network topologies are poised to break the Exaflops (10^18 floating point operations per second) barrier in the coming years. With the emergence of such highly hierarchical clusters, MPI has to be optimized for performance and scalability in order to cope with the ever-increasing demands of large-scale simulations. The proposed research is highly original and innovative in the sense that it addresses key issues in MPI, by including topology-awareness for process mapping, by incorporating dedicated queues for partner processes, by quality of service provisioning partner/non-partner traffic, and by developing fiber-based asynchronous progression techniques. The outcome of this research will be relevant to various sectors in Canada, including Environment Canada, Compute Canada, Canada Genome Sciences Centre, oil and gas industries, and ultimately the Canadian public at large. It is expected that the findings from this research will have significant impact on the target community, and that it will lead to new directions for future research. The proposed research is ideal for training HQP in that it has a strong foundation that translates immediately into practical applications and implementations. There is a high demand for graduates in HPC and networking, and the HQP trained will be well positioned to compete for jobs in academia and industry.