Summary Dataflow programming models have been growing in popularity as a means to deliver a good balance between performance and portability in the post‐petascale era. In this paper, we evaluate different dataflow programming models for electronic structure methods and compare them in terms of programmability, resource utilization, and scalability. In particular, we evaluate two programming paradigms for expressing scientific applications in a dataflow form: (1) explicit dataflow, where the dataflow is specified explicitly by the developer, and (2) implicit dataflow, where a task scheduling runtime derives the dataflow using per‐task data‐access information embedded in a serial program. We discuss our findings and present a thorough experimental analysis using methods from the NWChem quantum chemistry application as our case study, and OpenMP, StarPU, and PaRSEC as the task‐based runtimes that enable the different forms of dataflow execution. Furthermore, we derive an abstract model to explore the limits of the different dataflow programming paradigms.

Task-based execution has been growing in popularity as a means to deliver a good balance between performance and portability in the post-petascale era. The Parallel Runtime Scheduling and Execution Control (PARSEC) framework is a task-based runtime system that we designed to achieve high performance computing at scale. PARSEC offers a programming paradigm that is different than what has been traditionally used to develop large scale parallel scientific applications. In this paper, we discuss the use of PARSEC to convert a part of the Coupled Cluster (CC) component of the Quantum Chemistry package NWCHEM into a task-based form. We explain how we organized the computation of the CC methods in individual tasks with explicitly defined data dependencies between them and re-integrated the modified code into NWCHEM. We present a thorough performance evaluation and demonstrate that the modified code outperforms the original by more than a factor of two. We also compare the performance of different variants of the modified code and explain the different behaviors that lead to the differences in performance.

Computational chemistry comprises one of the driving forces of High Performance Computing. In particular, many-body methods, such as Coupled Cluster methods (CC) [1] of the quantum chemistry package NWCHEM [2], are of particular interest for the applied chemistry community.

For more than a decade, the PAPI performance-monitoring library has provided a clear, portable interface to the hardware performance counters available on all modern CPUs and other components of interest (e.g., GPUs, network, and I/O systems). Most major end-user tools that application developers use to analyze the performance of their applications rely on PAPI to gain access to these performance counters. One of the critical roadblocks on the way to larger, more complex high performance systems, has been widely identified as being the energy efficiency constraints. With modern extreme scale machines having hundreds of thousands of cores, the ability to reduce power consumption for each CPU at the software level becomes critically important, both for economic and environmental reasons. In order for PAPI to continue playing its well established role in HPC, it is pressing to provide valuable performance data that not only originates from within the processing cores but also delivers insight into the power consumption of the system as a whole. An extensive effort has been made to extend the Performance API to support power monitoring capabilities for various platforms. This paper provides detailed information about three components that allow power monitoring on the Intel Xeon Phi and Blue Gene/Q. Furthermore, we discuss the integration of PAPI in PARSEC - a task-based dataflow-driven execution engine - enabling hardware performance counter and power monitoring at true task granularity.

DARPA's AACE project aimed to develop Architecture Aware Compiler Environments. Such a compiler automatically characterizes the targeted hardware and optimizes the application codes accordingly. We present the BlackjackBench suite, a collection of portable micro-benchmarks that automate system characterization, plus statistical analysis techniques for interpreting the results. The BlackjackBench benchmarks discover the effective sizes and speeds of the hardware environment rather than the often unattainable peak values. We aim at hardware characteristics that can be observed by running executables generated by existing compilers from standard C codes. We characterize the memory hierarchy, including cache sharing and non-uniform memory access characteristics of the system, properties of the processing cores affecting the instruction execution speed and the length of the operating system scheduler time slot. We show how these features of modern multicores can be discovered programmatically. We also show how the features could potentially interfere with each other resulting in incorrect interpretation of the results, and how established classification and statistical analysis techniques can reduce experimental noise and aid automatic interpretation of results. We show how effective hardware metrics from our probes allow guided tuning of computational kernels that outperform an autotuning library further tuned by the hardware vendor.

Quality Threshold Clustering (QTC) is an algorithm for partitioning data, in fields such as biology, where clustering of large data-sets can aid scientific discovery. Unlike other clustering algorithms, QTC does not require knowing the number of clusters a priori, however, its perceived need for high computing power often makes it an unattractive choice. This paper presents a thorough study of QTC. We analyze the worst case complexity of the algorithm and discuss methods to reduce it by trading memory for computation. We also demonstrate how the expected running time of QTC is affected by the structure of the input data. We describe how QTC can be parallelized, and discuss implementation details of our thread-parallel, GPU, and distributed memory implementations of the algorithm. We demonstrate the efficiency of our implementations through experimental data. We show how data sets with tens of thousands of elements can be clustered in a matter of minutes in a modern GPU, and seconds in a small scale cluster of multi-core CPUs, or multiple GPUs. Finally, we discuss how user selected parameters, as well as algorithmic and implementation choices, affect performance.

Several existing compiler transformations can help improve communication-computation overlap in MPI applications. However, traditional compilers treat calls to the MPI library as a black box with unknown side effects and thus miss potential optimizations. This paper's contributions enable the development of an MPI-aware optimizing compiler that can perform transformations exploiting knowledge of MPI call effects to increase communication-computa-tion overlap. We formulate a set of data flow equations and rules to describe the side effects of key MPI functions so an MPI-aware compiler can automatically assess the safety of transformations. After categorizing existing compiler transformations based on their effect on the application code, we present an optimization algorithm that specifies when and how to apply these optimizing transformations to achieve improved communication-computation overlap. By manually applying the optimization algorithm to kernels extracted from HYCOM and the NAS benchmarks, we show that even when transforming these highly optimized codes, execution time can be decreased by an average of over 30%.

A non-trivial challenge in high performance, cluster computing is the communication overhead introduced by the cluster interconnect. A common strategy for addressing this challenge is the use of communication-computation overlapping. In this paper, we introduce Gravel, a portable communication library designed to inter-operate with MPI for improving communication-computation overlapping. Selected MPI calls are semi-automatically replaced by Gravel calls only in key locations in an application where performance is critical. Gravel separates the data transfers from the handshake messages, enabling application developers to utilize Remote Data Memory Access (RDMA) directly, whether the data exchange scheme of their application is pure one-sided, or a more traditional two-sided. The Gravel API is much simpler than existing low level libraries and details (i.e., pointers) are hidden from the application layer so Gravel is usable in FORTRAN applications. This paper presents an overview of Gravel.

This paper describes a source to source compilation tool for optimizing MPI-based parallel applications. This tool is able to automatically apply a "prepushing" transformation that causes MPI programs to aggressively send data as soon as it is available, thus improving communication-computation overlap and improving application performance. In this paper we present asphalt_transformer; the Open64-based component of our framework, ASPhALT, responsible for automatically performing the prepushing transformation. We also present an extensive study of the performance gains witnessed from automatically transformed codes. In particular, we demonstrate how different levels of aggregation affect the performance of parallel programs executing various computation kernels on different clusters. Furthermore, we discuss the differences in performance improvement between the hand-optimized and automatically optimized codes, as well as the effect of automation on time-to-solution.

Applications that execute on parallel clusters face scalability concerns due to the high communication overhead that is usually associated with such environments. Modern network technologies that support remote direct memory access (RDMA) can offer true zero copy communication and reduce communication overhead by overlapping it with computation. For this approach to be effective the parallel application using the cluster must be structured in a way that enables communication computation overlapping. Unfortunately, the trade-off between maintainability and performance often leads to a structure that prevents exploiting the potential for communication computation overlapping. This paper describes a source-to-source optimizing transformation that can be performed by an automatic (or semi-automatic) system in order to restructure MPI codes towards maximizing communication-computation overlapping.

This paper presents program transformations directed toward improving communication-computation overlap in parallel programs that use MPI’s collective operations. Our transformations target a wide variety of applications focusing on scientific codes with computation loops that exhibit limited dependence among iterations. We include guidance for developers for transforming an application code in order to exploit the communicationcomputation overlap available in the underlying cluster, as well as a discussion of the performance improvements achieved by our transformations. We present results from a detailed study of the effect of the problem and message size, level of communication-computation overlap, and amount of communication aggregation on runtime performance in a cluster environment based on an RDMA-enabled network. The targets of our study are two scientific codes written by domain scientists, but the applicability of our work extends far beyond the scope of these two applications.

Dynamic impact analysis based on whole path profiling of method calls and returns has been shown to provide more useful predictions of software change impacts than method-level static slicing and to avoid the overhead of expensive dependency analysis needed for dynamic slicing-based impact analysis. This work presents the design, implementation, and evaluation of an online approach to dynamic impact analysis as an extension to the DynamoRIO binary code modification system and to the Jikes Research Virtual Machine. Storage and postmortem analysis of program traces, even compressed, are avoided.