SCI Publications

As computer architectures are rapidly evolving (e.g. those designed for exascale), multiple portability frameworks have been developed to avoid new architecture-specific development and tuning. However, portability frameworks depend on compilers for auto-vectorization and may lack support for explicit vectorization on heterogeneous platforms. Alternatively, programmers can use intrinsics-based primitives to achieve more efficient vectorization, but the lack of a gpu back-end for these primitives makes such code non-portable. A unified, portable, Single Instruction Multiple Data (simd) primitive proposed in this work, allows intrinsics-based vectorization on cpus and many-core architectures such as Intel Knights Landing (knl), and also facilitates Single Instruction Multiple Threads (simt) based execution on gpus. This unified primitive, coupled with the Kokkos portability ecosystem, makes it possible to develop explicitly vectorized code, which is portable across heterogeneous platforms. The new simd primitive is used on different architectures to test the performance boost against hard-to-auto-vectorize baseline, to measure the overhead against efficiently vectroized baseline, and to evaluate the new feature called the \logical vector length" (lvl). The simd primitive provides portability across cpus and gpus without any performance degradation being observed experimentally.

High performance computing frameworks utilizing CPUs, Nvidia GPUs, and/or Intel Xeon Phis necessitate portable and scalable solutions for application developers. Nvidia GPUs in particular present numerous portability challenges with a different programming model, additional memory hierarchies, and partitioned execution units among streaming multiprocessors. This work presents modifications to the Uintah asynchronous many-task runtime and the Kokkos portability library to enable one single codebase for complex multiphysics applications to run across different architectures. Scalability and performance results are shown on multiple architectures for a globally coupled radiation heat transfer simulation, ranging from a single node to 16384 Titan compute nodes.

The Uintah computational framework is used for the parallel solution of partial differential equations on adaptive mesh refinement grids using modern supercomputers. Uintah is structured with an application layer and a separate runtime system. Uintah is based on a distributed directed acyclic graph (DAG) of computational tasks, with a task scheduler that efficiently schedules and executes these tasks on both CPU cores and on-node accelerators. The runtime system identifies task dependencies, creates a task graph prior to the execution of these tasks, automatically generates MPI message tags, and automatically performs halo transfers for simulation variables. Automating halo transfers in a heterogeneous environment poses significant challenges when tasks compute within a few milliseconds, as runtime overhead affects wall time execution, or when simulation variables require large halos spanning most or all of the computational domain, as task dependencies become expensive to process. These challenges are magnified at production scale when application developers require each compute node perform thousands of different halo transfers among thousands simulation variables. The principal contribution of this work is to (1) identify and address inefficiencies that arise when mapping tasks onto the GPU in the presence of automated halo transfers, (2) implement new schemes to reduce runtime system overhead, (3) minimize application developer involvement with the runtime, and (4) show overhead reduction results from these improvements.

We present our recent experience working to design parallel functional control-flow analysis (CFA) using an encoding in Datalog and underlying relational algebra implemented for SIMD coprocessors and supercomputers. Control-flow analysis statically models the possible propagations of data and control through a target program, finitely obtaining a bound on reachable expressions and environments and on possible return and argument values. We used Souffl´e, a parallel CPU-based Datalog implementation from Oracle labs, and worked toward a new MPI-based distributed hash join implementation and an extension of the GPU-based relational algebra library RedFox.

In this paper, we provide introductions to functional flow analysis, Datalog, MPI, and CUDA, explaining the total process we are working on to bring these components together in an analysis pipeline toward the end of scaling functional program analyses by extracting their intrinsic parallelism in a principled manner.

The University of Utah's Carbon Capture Multidisciplinary Simulation Center (CCMSC) is using the Uintah Computational Framework to predict performance of a 1000 MWe ultra-supercritical clean coal boiler. The center aims to utilize the Intel Xeon Phi-based DOE systems, Theta and Aurora, through the Aurora Early Science Program by using the Kokkos C++ library to enable node-level performance portability. This paper describes infrastructure advancements and portability improvements made possible by our integration of Kokkos within Uintah. Scalability results are presented that compare serial and data parallel task execution models for a challenging radiative heat transfer calculation, central to the center's predictive boiler simulations. These results demonstrate both good strong-scaling characteristics to 256 Knights Landing (KNL) processors on the NSF Stampede system, and show the KNL-based calculation to compete with prior GPU-based results for the same calculation.

Large-scale parallel applications with complex global data dependencies beyond those of reductions pose significant scalability challenges in an asynchronous runtime system. Internodal challenges include identifying the all-to-all communication of data dependencies among the nodes. Intranodal challenges include gathering together these data dependencies into usable data objects while avoiding data duplication. This paper addresses these challenges within the context of a large-scale, industrial coal boiler simulation using the Uintah asynchronous many-task runtime system on GPU architectures. We show significant reduction in time spent analyzing data dependencies through refinements in our dependency search algorithm. Multiple task graphs are used to eliminate subsequent analysis when task graphs change in predictable and repeatable ways. Using a combined data store and task scheduler redesign reduces data dependency duplication ensuring that problems fit within host and GPU memory. These modifications did not require any changes to application code or sweeping changes to the Uintah runtime system. We report results running on the DOE Titan system on 119K CPU cores and 7.5K GPUs simultaneously. Our solutions can be generalized to other task dependency problems with global dependencies among thousands of nodes which must be processed efficiently at large scale.

Background:
Image segmentation and registration techniques have enabled biologists to place large amounts of volume data from fluorescence microscopy, morphed three-dimensionally, onto a common spatial frame. Existing tools built on volume visualization pipelines for single channel or red-green-blue (RGB) channels have become inadequate for the new challenges of fluorescence microscopy. For a three-dimensional atlas of the insect nervous system, hundreds of volume channels are rendered simultaneously, whereas fluorescence intensity values from each channel need to be preserved for versatile adjustment and analysis. Although several existing tools have incorporated support of multichannel data using various strategies, the lack of a flexible design has made true many-channel visualization and analysis unavailable. The most common practice for many-channel volume data presentation is still converting and rendering pseudosurfaces, which are inaccurate for both qualitative and quantitative evaluations.

Results:
Here, we present an alternative design strategy that accommodates the visualization and analysis of about 100 volume channels, each of which can be interactively adjusted, selected, and segmented using freehand tools. Our multichannel visualization includes a multilevel streaming pipeline plus a triple-buffer compositing technique. Our method also preserves original fluorescence intensity values on graphics hardware, a crucial feature that allows graphics-processing-unit (GPU)-based processing for interactive data analysis, such as freehand segmentation. We have implemented the design strategies as a thorough restructuring of our original tool, FluoRender.

Conclusion:
The redesign of FluoRender not only maintains the existing multichannel capabilities for a greatly extended number of volume channels, but also enables new analysis functions for many-channel data from emerging biomedical-imaging techniques.

Modern supercomputers have thousands of nodes, each with CPUs and/or GPUs capable of several teraflops. However, the network connecting these nodes is relatively slow, on the order of gigabits per second. For time-critical workloads such as interactive visualization, the bottleneck is no longer computation but communication. In this paper, we present an image compositing algorithm that works on both CPU-only and GPU-accelerated supercomputers and focuses on communication avoidance and overlapping communication with computation at the expense of evenly balancing the workload. The algorithm has three stages: a parallel direct send stage, followed by a tree compositing stage and a gather stage. We compare our algorithm with radix-k and binary-swap from the IceT library in a hybrid OpenMP/MPI setting on the Stampede and Edison supercomputers, show strong scaling results and explain how we generally achieve better performance than these two algorithms. We developed a GPU-based image compositing algorithm where we use CUDA kernels for computation and GPU Direct RDMA for inter-node GPU communication. We tested the algorithm on the Piz Daint GPU-accelerated supercomputer and show that we achieve performance on par with CPUs. Lastly, we introduce a workflow in which both rendering and compositing are done on the GPU.

Isosurface extraction is a fundamental technique used for both surface reconstruction and mesh generation. One method to extract well-formed isosurfaces is a particle system; unfortunately, particle systems can be slow. In this paper, we introduce an enhanced parallel particle system that uses the closest point embedding as the surface representation to speedup the particle system for isosurface extraction. The closest point embedding is used in the Closest Point Method (CPM), a technique that uses a standard three dimensional numerical PDE solver on two dimensional embedded surfaces. To fully take advantage of the closest point embedding, it is coupled with a Barnes-Hut tree code on the GPU. This new technique produces well-formed, conformal unstructured triangular and tetrahedral meshes from labeled multi-material volume datasets. Further, this new parallel implementation of the particle system is faster than any known methods for conformal multi-material mesh extraction. The resulting speed-ups gained in this implementation can reduce the time from labeled data to mesh from hours to minutes and benefits users, such as bioengineers, who employ triangular and tetrahedral meshes.

Keywords: scalar field methods, GPGPU, curvature based, scientific visualization

Fast, scalable, low-cost, and low-power execution of parallel graph algorithms is important for a wide variety of commercial and public sector applications. Breadth First Search (BFS) imposes an extreme burden on memory bandwidth and network communications and has been proposed as a benchmark that may be used to evaluate current and future parallel computers. Hardware trends and manufacturing limits strongly imply that many core devices, such as NVIDIA® GPUs and the Intel® Xeon Phi®, will become central components of such future systems. GPUs are well known to deliver the highest FLOPS/watt and enjoy a very significant memory bandwidth advantage over CPU architectures. Recent work has demonstrated that GPUs can deliver high performance for parallel graph algorithms and, further, that it is possible to encapsulate that capability in a manner that hides the low level details of the GPU architecture and the CUDA language but preserves the high throughput of the GPU. We extend previous research on GPUs and on scalable graph processing on super-computers and demonstrate that a high-performance parallel graph machine can be created using commodity GPUs and networking hardware.

Keywords: GPU cluster, MPI, BFS, graph, parallel graph algorithm

Fast, scalable, low-cost, and low-power execution of parallel graph algorithms is important for a wide variety of commercial and public sector applications. Breadth First Search (BFS) imposes an extreme burden on memory bandwidth and network communications and has been proposed as a benchmark that may be used to evaluate current and future parallel computers. Hardware trends and manufacturing limits strongly imply that many core devices, such as NVIDIA® GPUs and the Intel® Xeon Phi®, will become central components of such future systems. GPUs are well known to deliver the highest FLOPS/watt and enjoy a very significant memory bandwidth advantage over CPU architectures. Recent work has demonstrated that GPUs can deliver high performance for parallel graph algorithms and, further, that it is possible to encapsulate that capability in a manner that hides the low level details of the GPU architecture and the CUDA language but preserves the high throughput of the GPU. We extend previous research on GPUs and on scalable graph processing on super-computers and demonstrate that a high-performance parallel graph machine can be created using commodity GPUs and networking hardware.

Keywords: GPU cluster, MPI, BFS, graph, parallel graph algorithm

Recent trends in high performance computing present larger and more diverse computers using multicore nodes possibly with accelerators and/or coprocessors and reduced memory. These changes pose formidable challenges for applications code to attain scalability. Software frameworks that execute machine-independent applications code using a runtime system that shields users from architectural complexities offer a portable solution for easy programming. The Uintah framework, for example, solves a broad class of large-scale problems on structured adaptive grids using fluid-flow solvers coupled with particle-based solids methods. However, the original Uintah code had limited scalability as tasks were run in a predefined order based solely on static analysis of the task graph and used only message passing interface (MPI) for parallelism. By using a new hybrid multithread and MPI runtime system, this research has made it possible for Uintah to scale to 700K central processing unit (CPU) cores when solving challenging fluid-structure interaction problems. Those problems often involve moving objects with adaptive mesh refinement and thus with highly variable and unpredictable work patterns. This research has also demonstrated an ability to run capability jobs on the heterogeneous systems with Nvidia graphics processing unit (GPU) accelerators or Intel Xeon Phi coprocessors. The new runtime system for Uintah executes directed acyclic graphs of computational tasks with a scalable asynchronous and dynamic runtime system for multicore CPUs and/or accelerators/coprocessors on a node. Uintah's clear separation between application and runtime code has led to scalability increases without significant changes to application code. This research concludes that the adaptive directed acyclic graph (DAG)-based approach provides a very powerful abstraction for solving challenging multiscale multiphysics engineering problems. Excellent scalability with regard to the different processors and communications performance are achieved on some of the largest and most powerful computers available today.

We present a novel algorithm for computing hierarchical geodesic models (HGMs) for diffeomorphic longitudinal shape analysis. The proposed algorithm exploits the inherent parallelism arising out of the independence in the contributions of individual geodesics to the group geodesic. The previous serial implementation severely limits the use of HGMs to very small population sizes due to computation time and massive memory requirements. The conventional method makes it impossible to estimate the parameters of HGMs on large datasets due to limited memory available onboard current GPU computing devices. The proposed parallel algorithm easily scales to solve HGMs on a large collection of 3D images of several individuals. We demonstrate its effectiveness on longitudinal datasets of synthetically generated shapes and 3D magnetic resonance brain images (MRI).

Keywords: LDDMM, HGM, Vector Momentum, Diffeomorphisms, Longitudinal Analysis

Volume rendering continues to be a critical method for analyzing large-scale scalar fields, in disciplines as diverse as biomedical engineering and computational fluid dynamics. Commodity desktop hardware has struggled to keep pace with data size increases, challenging modern visualization software to deliver responsive interactions for O(N³) algorithms such as volume rendering. We target the data type common in these domains: regularly-structured data.

In this work, we demonstrate that the major limitation of most volume rendering approaches is their inability to switch the data sampling rate (and thus data size) quickly. Using a volume renderer inspired by recent work, we demonstrate that the actual amount of visualizable data for a scene is typically bound considerably lower than the memory available on a commodity GPU. Our instrumented renderer is used to investigate design decisions typically swept under the rug in volume rendering literature. The renderer is freely available, with binaries for all major platforms as well as full source code, to encourage reproduction and comparison with future research.

Visualizing large molecular data requires efficient means of rendering millions of data elements that combine glyphs, geometry and volumetric techniques. The geometric and volumetric loads challenge traditional rasterization-based vis methods. Ray casting presents a scalable and memory- efficient alternative, but modern techniques typically rely on GPU-based acceleration to achieve interactive rendering rates. In this paper, we present bnsView, a molecular visualization ray tracing framework that delivers fast volume rendering and ball-and-stick ray casting on both multi-core CPUs andmany-core Intel ® Xeon Phi^TM co-processors, implemented in a SPMD language that generates efficient SIMD vector code for multiple platforms without source modification. We show that our approach running on co- processors is competitive with similar techniques running on GPU accelerators, and we demonstrate large-scale parallel remote visualization from TACC's Stampede supercomputer to large-format display walls using this system.

In this work, we describe our preliminary experiences on the Stampede system in the context of the Uintah Computational Framework. Uintah was developed to provide an environment for solving a broad class of fluid-structure interaction problems on structured adaptive grids. Uintah uses a combination of fluid-flow solvers and particle-based methods, together with a novel asynchronous taskbased approach and fully automated load balancing. While we have designed scalable Uintah runtime systems for large CPU core counts, the emergence of heterogeneous systems presents considerable challenges in terms of effectively utilizing additional on-node accelerators and co-processors, deep memory hierarchies, as well as managing multiple levels of parallelism. Our recent work has addressed the emergence of heterogeneous CPU/GPU systems with the design of a Unified heterogeneous runtime system, enabling Uintah to fully exploit these architectures with support for asynchronous, out-of-order scheduling of both CPU and GPU computational tasks. Using this design, Uintah has run at full scale on the Keeneland System and TitanDev. With the release of the Intel Xeon Phi co-processor and the recent availability of the Stampede system, we show that Uintah may be modified to utilize such a coprocessor based system. We also explore the different usage models provided by the Xeon Phi with the aim of understanding portability of a general purpose framework like Uintah to this architecture. These usage models range from the pragma based offload model to the more complex symmetric model, utilizing all co-processor and host CPU cores simultaneously. We provide preliminary results of the various usage models for a challenging adaptive mesh refinement problem, as well as a detailed account of our experience adapting Uintah to run on the Stampede system. Our conclusion is that while the Stampede system is easy to use, obtaining high performance from the Xeon Phi co-processors requires a substantial but different investment to that needed for GPU-based systems.

Keywords: Uintah, hybrid parallelism, scalability, parallel, adaptive, MIC, Xeon Phi, heterogeneous systems, Stampede, co-processor

Present trends in high performance computing present formidable challenges for applications code using multicore nodes possibly with accelerators and/or co-processors and reduced memory while still attaining scalability. Software frameworks that execute machine-independent applications code using a runtime system that shields users from architectural complexities offer a possible solution. The Uintah framework for example, solves a broad class of large-scale problems on structured adaptive grids using fluid-flow solvers coupled with particle-based solids methods. Uintah executes directed acyclic graphs of computational tasks with a scalable asynchronous and dynamic runtime system for CPU cores and/or accelerators/co-processors on a node. Uintah's clear separation between application and runtime code has led to scalability increases of 1000x without significant changes to application code. This methodology is tested on three leading Top500 machines; OLCF Titan, TACC Stampede and ALCF Mira using three diverse and challenging applications problems. This investigation of scalability with regard to the different processors and communications performance leads to the overall conclusion that the adaptive DAG-based approach provides a very powerful abstraction for solving challenging multi-scale multi-physics engineering problems on some of the largest and most powerful computers available today.

Keywords: Blue Gene/Q, GPU, Xeon Phi, adaptive, application, co-processor, heterogeneous systems, hybrid parallelism, parallel, scalability, software, uintah, NETL

In this work, we describe our preliminary experiences on the Stampede system in the context of the Uintah Computational Framework. Uintah was developed to provide an environment for solving a broad class of fluid-structure interaction problems on structured adaptive grids. Uintah uses a combination of fluid-flow solvers and particle-based methods, together with a novel asynchronous task-based approach and fully automated load balancing. While we have designed scalable Uintah runtime systems for large CPU core counts, the emergence of heterogeneous systems presents considerable challenges in terms of effectively utilizing additional on-node accelerators and co-processors, deep memory hierarchies, as well as managing multiple levels of parallelism. Our recent work has addressed the emergence of heterogeneous CPU/GPU systems with the design of a Unified heterogeneous runtime system, enabling Uintah to fully exploit these architectures with support for asynchronous, out-of-order scheduling of both CPU and GPU computational tasks. Using this design, Uintah has run at full scale on the Keeneland System and TitanDev. With the release of the Intel Xeon Phi co-processor and the recent availability of the Stampede system, we show that Uintah may be modified to utilize such a co-processor based system. We also explore the different usage models provided by the Xeon Phi with the aim of understanding portability of a general purpose framework like Uintah to this architecture. These usage models range from the pragma based offload model to the more complex symmetric model, utilizing all co-processor and host CPU cores simultaneously. We provide preliminary results of the various usage models for a challenging adaptive mesh refinement problem, as well as a detailed account of our experience adapting Uintah to run on the Stampede system. Our conclusion is that while the Stampede system is easy to use, obtaining high performance from the Xeon Phi co-processors requires a substantial but different investment to that needed for GPU-based systems.

Keywords: MIC, Xeon Phi, adaptive, co-processor, heterogeneous systems, hybrid parallelism, parallel, scalability, stampede, uintah, c-safe

Brainbow is a genetic engineering technique that randomly colorizes cells. Biological samples processed with this technique and imaged with confocal microscopy have distinctive colors for individual cells. Complex cellular structures can then be easily visualized. However, the complexity of the Brainbow technique limits its applications. In practice, most confocal microscopy scans use different florescence staining with typically at most three distinct cellular structures. These structures are often packed and obscure each other in rendered images making analysis difficult. In this paper, we leverage a process known as GPU framebuffer feedback loops to synthesize Brainbow-like images. In addition, we incorporate ID shuffling and Monte-Carlo sampling into our technique, so that it can be applied to single-channel confocal microscopy data. The synthesized Brainbow images are presented to domain experts with positive feedback. A user survey demonstrates that our synthetic Brainbow technique improves visualizations of volume data with complex structures for biologists.