Achievements

Exascale supercomputers will deliver 10^18 floating-point operations per second (FLOPS) in double precision using the High Performance Linpack (HPL) benchmark. Today, the two fastest supercomputers are the Chinese Sunway TaihuLight and Tianhe-2 supercomputers delivering respectively 93 and 33.8 petaFLOPS. The next American supercomputers from the Corral initiative will deliver between 100 and 200 petaFLOPS. Current projections of future supercomputers estimate the delivery of exascale machine in 2024.

While the race to exascale resulted in faster and larger supercomputers that today are only a factor of ten far from exascale, the software stack to support parallel applications on supercomputers has remained almost unchanged with respect to the software present on petascale machines. In fact, almost all applications use MPI as main library to support parallel communication. Some of the applications use Partitioned Global Address Space (PGAS) languages, such as Coarray Fortran and UPC, with MPI. Several applications employ OpenMP in combination with MPI for programming intra-node communication. Despite the proposal of new disruptive programming models and systems [1, 3, 5], it is unlikely that such programming systems will reach a level of maturity and reliability to be readily deployed on the full exascale machine. The programming system dominating the exascale era will be MPI, possibly in combination with PGAS and OpenMP. For this reason, it is important both to improve the performance of these existing programming systems and to enhance their interoperability.

The size of next exascale supercomputers and their hardware pose difficult challenges to development of programming models. One of the main challenges is to handle the amount of parallelism that an exascale supercomputer will provide. The current fastest supercomputer Sunway TaihuLight provides 10,649,600 cores for parallel computation. Extrapolating this value to future more powerful supercomputers, it is reasonable to expect that an exascale machine will provide an 100 million-way parallelism. While communication cost decreases as part of communication operations is offloaded to the NIC and network technologies improve, still a large amount of memory is needed for storing process and communicator information on 100 million processes and memory footprint becomes a serious bottleneck. In addition, collective operations and synchronization of such a large amount of processes require the development and implementation of more sophisticated collective algorithms. A second main challenge is to guarantee that all the programming systems, such MPI, OpenMP and PGAS approaches efficiently interoperate sharing fairly all the hardware resources. It is therefore important to improve the performance of communication operations on a very large number of process potentially in presence of a combination of different programming systems.

Exascale ProGRAmming Models (EPiGRAM) is a European Commission funded project with the goal of addressing these exascale challenges in programming models. The EPiGRAM consortium consists of KTH Royal Institute of Technology, Vienna University of Technology (TU Wien), Fraunhofer ITWM, Cray UK, University of Edinburgh and University of Illinois (associate partner).

EPiGRAM focuses on the improvement of MPI and GASPI performance on exascale systems. MPI is currently the most used approach for programming parallel applications on supercomputers. Global Address Space Programming Interface (GASPI) is the standard for a PGAS API. GASPI uses one-sided Remote Direct Memory Access (RDMA) driven communication in combination with remote completion in a PGAS environment. Global address space Programming Interface (GPI) is a GASPI implementation, developed by Fraunhofer ITWM. Since GPI-2, Fraunhofer ITWM provides an open-source GPI implementation under GPL v3.

Figure 1 provides an overview of the different topics that have been investigated during the EPiGRAM project and the remaining part of the paper will describe these topics more in detail.

Overview of the EPiGRAM project

Fig. 1. Overview of the EPiGRAM project. The second section presents the EPiGRAM work to address the challenge of exascale parallelism. In particular, we present how MPI and GASPI are further developed to reduce the memory footprint, to improve collective algorithms and implementations and to support emerging computing models. The third section describes the EPiGRAM research on MPI and GASPI interoperability, presenting MPI endpoints, GPI interoperability with MPI and the PGAS-based MPI implementation, called EMPI4Re, that integrates message-passing and PGAS in one framework. The fourth section presents the EPiGRAM applications. Finally, the fifth section concludes the paper summarizing the work done in EPiGRAM.

Highlighted publications