Preparing HPC Applications for the Exascale Era: A Decoupling Strategy

Abstract

Production-quality parallel applications are often a mixture of diverse operations, such as computation-and communication-intensive, regular and irregular, tightly coupled and loosely linked operations. In conventional construction of parallel applications, each process performs all the operations, which might result inefficient and seriously limit scalability, especially at large scale. We propose a decoupling strategy to improve the scalability of applications running on large-scale systems. Our strategy separates application operations onto groups of processes and enables a dataflow processing paradigm among the groups. This mechanism is effective in reducing the impact of load imbalance and increases the parallel efficiency by pipelining multiple operations. We provide a proof-of-concept implementation using MPI, the de-facto programming system on current supercomputers. We demonstrate the effectiveness of this strategy by decoupling the reduce, particle communication, halo exchange and I/O operations in a set of scientific and dataanalytics applications. A performance evaluation on 8,192 processes of a Cray XC40 supercomputer shows that the proposed approach can achieve up to 4× performance improvement.

SAGE: Percipient Storage for Exascale Data Centric Computing

“…We are starting to quantify the benefits of the individual features sets of the SAGE stack in the ongoing process of providing a holistic picture of the benefits of the SAGE architecture for Exascale. On that front, we present the performance results of two components of the high-level HPC interfaces for the SAGE platform: the PGAS I/O using MPI storage windows [30] and MPI streams [32,31,16] for post-processing and efficient parallelI/O operations, exploiting the SAGE architecture. This aims to provide introductory results for SAGE stack components with results for other components targeted for future works.…”

“…In particular, our work in SAGE focuses on parallel streams, where data producers and consumers are distributed among processes that require communication to move data. To achieve this, we have developed a stream library, called MPIStream, to support post-processing and parallel I/O operations on MPI consumer processes [31,16]. More details about MPI streams operation semantics and MP-IStream implementation are provided in Ref.…”

“…The ability to run in-storage compute is crucial to avoid the energy costs incurred by constantly moving data between the compute and the storage system [15]. Within SAGE, we use the function-shipping capability of the Mero object storage platform to compute in-storage and use MPI streams for offloading computations to processes running on I/O servers [16].…”

Decoupled Strategy for Imbalanced Workloads in MapReduce Frameworks

“…Given the irregular nature of certain input datasets, these design considerations pose performance restrictions when the workload per process becomes unexpectedly unbalanced. In such cases, it has been demonstrated that the use of decentralized algorithms can provide significant performance benefits [14].…”