SHF: Large: Collaborative Research: Unified Runtime for Supporting Hybrid Programming Models on Heterogeneous Architecture

点击翻译按钮获取中文摘要

Most of the traditional High-End Computing (HEC) applications andcurrent petascale applications are written using the Message PassingInterface (MPI) programming model. Some of these applications are runin MPI+OpenMP mode. However, it can be very difficult to use MPI orMPI+OpenMP and maintain performance for applications which demonstrateirregular and dynamic communication patterns. The Partitioned GlobalAddress Space (PGAS) programming model presents a flexible way forthese applications to express parallelism. Accelerators introduceadditional programming models: CUDA, OpenCL or OpenACC. Thus, theemerging heterogeneous architectures require support for varioushybrid programming models: MPI+OpenMP, MPI+PGAS, and MPI+PGAS+OpenMPwith extended APIs for multiple levels of parallelism. Unfortunately,there is no unified runtime which delivers the best performance andscalability for all of these hybrid programming models for a range ofapplications on current and next-generation HEC systems. This leadsto the following broad challenge: "Can a unified runtime for hybridprogramming model be designed which can provide benefits that aregreater than the sum of its parts?"A synergistic and comprehensive research plan, involving computerscientists from The Ohio State University (OSU) and Ohio SupercomputerCenter (OSC) and computational scientists from the Texas AdvancedComputing Center (TACC) and San Diego Supercomputer Center (SDSC),University of California San Diego (UCSD), is proposed to address theabove broad challenge with innovative solutions. The investigatorswill specifically address the following challenges: 1) What are therequirements and limitations of using hybrid programming models for aset of petascale applications? 2) What features and mechanisms areneeded in a unified runtime? 3) How can the unified runtime andassociated extension to programming model APIs be designed andimplemented? 4) How can candidate petascale applications beredesigned to take advantage of proposed unified runtime? and 5) Whatkind of benefits (in terms of performance, scalability andproductivity) can be achieved by the proposed approach? The researchwill be driven by a set of applications from established NSFcomputational science researchers running large scale simulations onRanger and other systems at OSC, SDSC and OSU. The proposed designswill be integrated into the open-source MVAPICH2 library. Theestablished national-scale training and outreach programs at TACC,SDSC and OSC will be used to disseminate the results of this research.