Collaborative Research: Extending One-Sided Communication in MPI Programming Model for Next-Generation Ultra-Scale HEC

大多数传统的HEC应用程序和当前的千万亿级应用程序都是使用消息传递接口(MPI)编程模型编写的。MPI-1标准为双边操作提供了通信语义。MPI-2标准增加了新的单向通信语义。然而，目前大多数候选的Petascale应用程序仍然使用MPI-1语义，这些应用程序发现现有MPI-2库中可用的MPI单边通信语义及其实现对开发性能、可伸缩性和容错能力具有很大的限制。调查人员包括来自俄亥俄州立大学(OSU)的计算机科学家以及来自德克萨斯高级计算中心(TACC)和圣地亚哥超级计算机中心(SDSC)的计算科学家，他们将研究和分析MPI单边通信语义、其实现和用法中当前的限制。人们将提出新的解决方案来缓解这些限制，从而使下一代超大规模系统和应用程序能够扩展到数十万个内核。研究人员将具体解决以下挑战：1)在千万亿级应用程序中使用MPI单边操作的限制是什么？2)可以对当前的MPI单侧操作进行哪些扩展来缓解这些限制？3)如何在MPI库中为新兴的超大规模HEC系统设计和实现这些扩展？4)如何重新设计千兆级应用程序以利用所提出的单侧扩展及其实现？以及5)在下一代超大规模系统上为千万亿级应用程序建议的扩展可以实现什么样的好处(性能、可扩展性和容错)？这项研究将由一组Petascale应用程序(Enzo、AWM-Olsen、PSDNS和MPCUGLES)推动，这些应用程序来自NSF计算科学研究人员，他们在TACC Ranger和其他NSF HEC系统上运行大规模模拟。

Most of the traditional HEC applications and current petascale applications are written using the Message Passing Interface (MPI) programming model. The MPI-1 standard provides communication semantics for two-sided operations. The MPI-2 standard added new one-sided communication semantics. However, most of the current candidate petascale applications continue to use the MPI-1 semantics.These applications find the available MPI one-sided communication semantics and their implementations in existing MPI-2 libraries very restrictive to exploit performance, scalability and fault-tolerance. The investigators, involving computer scientists from Ohio State University (OSU) and computational scientists from Texas Advanced Computing Center (TACC) and San Diego Supercomputer Center (SDSC), will study and analyze the current restrictions in the MPI one-sided communication semantics, their implementations and usages. Novel solutions will be proposed to alleviate these restrictions so that the next generation ultra-scale systems and applications can be scaled to hundreds of thousands of cores.The investigators will specifically address the following challenges: 1) What are the limitations of using MPI one-sided operations in petascale applications? 2) What extensions are possible to the current MPI one-sided operations to alleviate such limitations? 3) How to design and implement these extensions in an MPI library for emerging ultra-scale HEC systems? 4) How to redesign petascale applications to take advantage of proposed one-sided extensions and their implementations? and 5) What kind of benefits (performance, scalability and fault tolerance) can be achieved by the proposed extensions for petascale applications on the next generation ultra-scale systems? The research will be driven by a set of petascale applications (ENZO, AWM-Olsen, PSDNS and MPCUGLES) from established NSF computational science researchers running large scale simulations on the TACC Ranger and other NSF HEC systems.