Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “MPI message data”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 19 records
All sourcesNASA NTRSDOE OSTI
DOE OSTI2021

Evaluating MPI resource usage summary statistics

The Message Passing Interface (MPI) remains the dominant programming model for scientific applications running on today’s high-performance computing (HPC) systems. This dominance stems from MPI’s powerful semantics for inter-process communication that has enabled scientists to write applications for simulating important physical phenomena. MPI does not, however, specify how messages and synchronization should be carried out. Those details are typically dependent on low-level architecture details and the message characteristics of the application. Therefore, analyzing an application’s MPI resource usage is critical to tuning MPI’s performance on a particular platform. The result of this analysis is typically a discussion of the mean message sizes, queue search lengths and message arrival times for a workload or set of workloads. While a discussion of the arithmetic mean in MPI resource usage might be the most intuitive summary statistic, it is not always the most accurate in terms of representing the underlying data. In this paper, we analyze MPI resource usage for a number of key MPI workloads using an existing MPI trace collector and discrete-event simulator. Our analysis demonstrates that the average, while easy and efficient to calculate, is a useful metric for characterizing latency and bandwidth measurements, but may not be a good representation of application message sizes, match list search depths, or MPI inter-operation times. Additionally, we show that the median and mode are superior choices in many cases. We also observe that the arithmetic mean is not the best representation of central tendency for data that are drawn from distributions that are multi-modal or have heavy tails. Furthermore, the results and analysis of our work provide valuable guidance on how we, as a community, should discuss and analyze MPI resource usage data for scientific applications.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

LowFive v1.0

LowFive is a new data transport layer based on the HDF5 data model, for in situ workflows. Executables using LowFive can communicate in situ (using in-memory data and MPI message passing), reading and writing traditional HDF5 files to physical storage, and combining the two modes. Minimal and often no source-code modification is needed for programs that already use HDF5. LowFive maintains deep copies or shallow references of datasets, configurable by the user. More than one task can produce (write) data, and more than one task can consume (read) data, accommodating fan-in and fan-out in the workflow task graph. LowFive supports data redistribution from n producer processes to m consumer processes.

Morozov, Dmitriy↗
DOE OSTI2024

Implementing One Sided Partitioned Communication in Open MPI

This report introduces partitioned communication, a new MPI 4.0 interface that enables early bird communication by overlapping communication and computation. By partitioning messages into smaller sub-messages, MPI can start partial data transfers early. Performance studies show that the RMA implementation outperforms the Persistent implementation, despite some constraints. This report details a new opt-in RMA implementation, offering a high-performance option for partitioned communication that imposes some additional limitations.

97 MATHEMATICS AND COMPUTING↗
NASA NTRS2000

Performance Assessment of OVERFLOW on Distributed Computing Environment

The aerodynamic computer code, OVERFLOW, with a multi-zone overset grid feature, has been parallelized to enhance its performance on distributed and shared memory paradigms. Practical application benchmarks have been set to assess the efficiency of code's parallelism on high-performance architectures. The code's performance has also been experimented with in the context of the distributed computing paradigm on distant computer resources using the Information Power Grid (IPG) toolkit, Globus. Two parallel versions of the code, namely OVERFLOW-MPI and -MLP, have developed around the natural coarse grained parallelism inherent in a multi-zonal domain decomposition paradigm. The algorithm invokes a strategy that forms a number of groups, each consisting of a zone, a cluster of zones and/or a partition of a large zone. Each group can be thought of as a process with one or multithreads assigned to it and that all groups run in parallel. The -MPI version of the code uses explicit message-passing based on the standard MPI library for sending and receiving interzonal boundary data across processors. The -MLP version employs no message-passing paradigm; the boundary data is transferred through the shared memory. The -MPI code is suited for both distributed and shared memory architectures, while the -MLP code can only be used on shared memory platforms. The IPG applications are implemented by the -MPI code using the Globus toolkit. While a computational task is distributed across multiple computer resources, the parallelism can be explored on each resource alone. Performance studies are achieved with some practical aerodynamic problems with complex geometries, consisting of 2.5 up to 33 million grid points and a large number of zonal blocks. The computations were executed primarily on SGI Origin 2000 multiprocessors and on the Cray T3E. OVERFLOW's IPG applications are carried out on NASA homogeneous metacomputing machines located at three sites, Ames, Langley and Glenn. Plans for the future will exploit the distributed parallel computing capability on various homogeneous and heterogeneous resources and large scale benchmarks. Alternative IPG toolkits will be used along with sophisticated zonal grouping strategies to minimize the communication time across the computer resources.

Djomehri, M. Jahed↗
NASA NTRS2004

What Multilevel Parallel Programs do when you are not Watching: A Performance Analysis Case Study Comparing MPI/OpenMP, MLP, and Nested OpenMP

With the current trend in parallel computer architectures towards clusters of shared memory symmetric multi-processors, parallel programming techniques have evolved that support parallelism beyond a single level. When comparing the performance of applications based on different programming paradigms, it is important to differentiate between the influence of the programming model itself and other factors, such as implementation specific behavior of the operating system (OS) or architectural issues. Rewriting-a large scientific application in order to employ a new programming paradigms is usually a time consuming and error prone task. Before embarking on such an endeavor it is important to determine that there is really a gain that would not be possible with the current implementation. A detailed performance analysis is crucial to clarify these issues. The multilevel programming paradigms considered in this study are hybrid MPI/OpenMP, MLP, and nested OpenMP. The hybrid MPI/OpenMP approach is based on using MPI [7] for the coarse grained parallelization and OpenMP [9] for fine grained loop level parallelism. The MPI programming paradigm assumes a private address space for each process. Data is transferred by explicitly exchanging messages via calls to the MPI library. This model was originally designed for distributed memory architectures but is also suitable for shared memory systems. The second paradigm under consideration is MLP which was developed by Taft. The approach is similar to MPi/OpenMP, using a mix of coarse grain process level parallelization and loop level OpenMP parallelization. As it is the case with MPI, a private address space is assumed for each process. The MLP approach was developed for ccNUMA architectures and explicitly takes advantage of the availability of shared memory. A shared memory arena which is accessible by all processes is required. Communication is done by reading from and writing to the shared memory.

Jost, Gabriele↗
NASA NTRS2011

High-Performance Data Analysis Tools for Sun-Earth Connection Missions

The data analysis tool of choice for many Sun-Earth Connection missions is the Interactive Data Language (IDL) by ITT VIS. The increasing amount of data produced by these missions and the increasing complexity of image processing algorithms requires access to higher computing power. Parallel computing is a cost-effective way to increase the speed of computation, but algorithms oftentimes have to be modified to take advantage of parallel systems. Enhancing IDL to work on clusters gives scientists access to increased performance in a familiar programming environment. The goal of this project was to enable IDL applications to benefit from both computing clusters as well as graphics processing units (GPUs) for accelerating data analysis tasks. The tool suite developed in this project enables scientists now to solve demanding data analysis problems in IDL that previously required specialized software, and it allows them to be solved orders of magnitude faster than on conventional PCs. The tool suite consists of three components: (1) TaskDL, a software tool that simplifies the creation and management of task farms, collections of tasks that can be processed independently and require only small amounts of data communication; (2) mpiDL, a tool that allows IDL developers to use the Message Passing Interface (MPI) inside IDL for problems that require large amounts of data to be exchanged among multiple processors; and (3) GPULib, a tool that simplifies the use of GPUs as mathematical coprocessors from within IDL. mpiDL is unique in its support for the full MPI standard and its support of a broad range of MPI implementations. GPULib is unique in enabling users to take advantage of an inexpensive piece of hardware, possibly already installed in their computer, and achieve orders of magnitude faster execution time for numerically complex algorithms. TaskDL enables the simple setup and management of task farms on compute clusters. The products developed in this project have the potential to interact, so one can build a cluster of PCs, each equipped with a GPU, and use mpiDL to communicate between the nodes and GPULib to accelerate the computations on each node.

Messmer, Peter↗
NASA NTRS2022

VULCAN-CFD User Manual: Ver. 7.2.0

VULCAN-CFD offers a comprehensive set of capabilities to enable the simulation of continuum flowfields from subsonic to hypersonic conditions. The governing equations that are employed include allowances for both chemical and thermal nonequilibrium processes, coupled with a wide variety of turbulence models for both Reynolds-averaged and large eddy simulations. The software package can simulate two-dimensional, axisymmetric, or three-dimensional problems on structured multiblock meshes or fully unstructured meshes. A parabolic (i.e., space-marching) treatment can also be used for any subset of a structured mesh that can accommodate this solution strategy. The flow solver provides a significant level of geometric flexibility for structured grid simulations by allowing for arbitrary face-to-face C(0) continuous and non-C(0) continuous block interface connectivities. The unstructured grid paradigm allows for mixed element unstructured meshes that contain any combination of tetrahedral, prismatic, pyramidal, and hexahedral cell elements. The flow solver is also fully parallelized using MPI (Message Passing Interface) libraries in a data-parallel fashion, allowing for efficient simulations on modern High Performance Computing (HPC) systems. This document provides information related to the installation and execution of the VULCAN-CFD software package. A detailed description of the physical and numerical models available in the software are provided in the VULCAN-CFD Theory Manual.

VULCAN-CFD User Manual↗
NASA NTRS2019

Using Big Data Technologies with Earth Science Data in HDF5: HDF5 Scalable Solutions

HDF5 (Hierarchical Data Format 5) is open-source, high-performance software that consists of an abstract data model, library, and fileformat used for storing and managing extremely large and/or complex data collections. NASA Earth Observing System (EOS) Data and Information Systems use HDF5 as an archival format to store remote sensing data from EOS satellites. HDF5 is also used to store other types of Geoscience and Strophysical data, e.g., seismic data and data from Low-Frequency Array (LOFAR) radio telescopes. Data stored in HDF5 has reached tens of petabytes and is growing at an accelerated rate.With the growing amout of HDF5 Earth Science data to analyze and process, scientists need to adopt big data technologies including new storage paradigms such as cloud and object storage. To run models and perform data analysis they also need to utilizied efficient and diverse ways to access data, from high-performance computing's (HPC) Message Passing Interface (MPI) I/O and deep memory hierarchies (DMH) to non-HPC frameworks such as Apache Hadoop, Spark, and Drill. The HDF Group continually works to enable usage of big data technologies in HDF software.

Knox, Larry↗
DOE OSTI2024

MPI nuts and bolts and more [Slides]

MPI (Message-Passing Interface) is a message-passing library interface specification. All parts of this definition are significant. MPI addresses primarily the message-passing parallel programming model, in which data is moved from the address space of one process to that of another process through cooperative operations on each process. . . MPI is a specification, not an implementation; there are multiple implementations of MPI. This specification is for a library interface; MPI is not a language, and all MPI operations are expressed as functions, subroutines, or methods, according to the appropriate language bindings that, for C and Fortran, are part of the MPI standard. MPI Forum is the organization which is responsible for the MPI Specification.

97 MATHEMATICS AND COMPUTING↗
NASA NTRS1996

Parallel Simulation of Unsteady Turbulent Flames

Time-accurate simulation of turbulent flames in high Reynolds number flows is a challenging task since both fluid dynamics and combustion must be modeled accurately. To numerically simulate this phenomenon, very large computer resources (both time and memory) are required. Although current vector supercomputers are capable of providing adequate resources for simulations of this nature, the high cost and their limited availability, makes practical use of such machines less than satisfactory. At the same time, the explicit time integration algorithms used in unsteady flow simulations often possess a very high degree of parallelism, making them very amenable to efficient implementation on large-scale parallel computers. Under these circumstances, distributed memory parallel computers offer an excellent near-term solution for greatly increased computational speed and memory, at a cost that may render the unsteady simulations of the type discussed above more feasible and affordable.This paper discusses the study of unsteady turbulent flames using a simulation algorithm that is capable of retaining high parallel efficiency on distributed memory parallel architectures. Numerical studies are carried out using large-eddy simulation (LES). In LES, the scales larger than the grid are computed using a time- and space-accurate scheme, while the unresolved small scales are modeled using eddy viscosity based subgrid models. This is acceptable for the moment/energy closure since the small scales primarily provide a dissipative mechanism for the energy transferred from the large scales. However, for combustion to occur, the species must first undergo mixing at the small scales and then come into molecular contact. Therefore, global models cannot be used. Recently, a new model for turbulent combustion was developed, in which the combustion is modeled, within the subgrid (small-scales) using a methodology that simulates the mixing and the molecular transport and the chemical kinetics within each LES grid cell. Finite-rate kinetics can be included without any closure and this approach actually provides a means to predict the turbulent rates and the turbulent flame speed. The subgrid combustion model requires resolution of the local time scales associated with small-scale mixing, molecular diffusion and chemical kinetics and, therefore, within each grid cell, a significant amount of computations must be carried out before the large-scale (LES resolved) effects are incorporated. Therefore, this approach is uniquely suited for parallel processing and has been implemented on various systems such as: Intel Paragon, IBM SP-2, Cray T3D and SGI Power Challenge (PC) using the system independent Message Passing Interface (MPI) compiler. In this paper, timing data on these machines is reported along with some characteristic results.

Menon, Suresh↗
DOE OSTI2024

TeMPI Shim

This is the manual for the TeMPI Shim library, whose goal is to facilitate scientific apps being loosely coupled through MPI. TeMPI Shim can be used to create and manage Message Passing Interface (MPI, see [MPI]) Communicators for Single Program, Multiple Data (SPMD, see [SPMD]) and Multiple Programs, Multiple Data (MPMD, see [MPMD]) workflows. MPI provides the MPI_APPNUM command number to each of the processes within its context. This number, starting at 0, is the application number being executed. In a case where 3 applications are being run in MPMD mode, there would be MPI_APPNUM values of 0, 1, and 2. TeMPI Shim creates intra- and inter-communicators between each pair of MPI_APPNUM values. In the aforementioned case, application 0 would have intra- and inter-communicators to speak to itself, application 1, and application 2. This is replicated for each of the applications. Additionally, TeMPI Shim creates intra- and inter-communicators for the first MPI rank of each application to directly communicate only with each other. Finally, TeMPI Shim creates its own copy of the default world communicator, i.e., MPI_COMM_WORLD. In the case where there is only a single application, it will have the communicators to only speak with itself. TeMPI Shim is useful in this case since it is considered good practice for MPI applications to copy the default world communicator and reference this copy (see [Duplicate World]_) anyways. Ultimately, it provides value independent of the number of applications present.

96 KNOWLEDGE MANAGEMENT AND PRESERVATION↗
NASA NTRS1995

RISC Processors and High Performance Computing

In this tutorial, we will discuss top five current RISC microprocessors: The IBM Power2, which is used in the IBM RS6000/590 workstation and in the IBM SP2 parallel supercomputer, the DEC Alpha, which is in the DEC Alpha workstation and in the Cray T3D; the MIPS R8000, which is used in the SGI Power Challenge; the HP PA-RISC 7100, which is used in the HP 700 series workstations and in the Convex Exemplar; and the Cray proprietary processor, which is used in the new Cray J916. The architecture of these microprocessors will first be presented. The effective performance of these processors will then be compared, both by citing standard benchmarks and also in the context of implementing a real applications. In the process, different programming models such as data parallel (CM Fortran and HPF) and message passing (PVM and MPI) will be introduced and compared. The latest NAS Parallel Benchmark (NPB) absolute performance and performance per dollar figures will be presented. The next generation of the NP13 will also be described. The tutorial will conclude with a discussion of general trends in the field of high performance computing, including likely future developments in hardware and software technology, and the relative roles of vector supercomputers tightly coupled parallel computers, and clusters of workstations. This tutorial will provide a unique cross-machine comparison not available elsewhere.

Saini, Subhash↗
DOE OSTI2021

TEMPI: An Interposed MPI Library with Canonical Representation of MPI Datatypes [Slides]

These points are covered in this presentation: Distributed GPU stencil, non-contiguous data; Equivalence of strided datatypes and minimal representation; GPU communication methods; Deploying on managed systems; Large messages and MPI datatypes; Translation and canonicalization; Automatic model-driven transfer method selection; and Interposed library implementation.

97 MATHEMATICS AND COMPUTING↗
NASA NTRS1997

Cooperative Data Sharing: Simple Support for Clusters of SMP Nodes

Libraries like PVM and MPI send typed messages to allow for heterogeneous cluster computing. Lower-level libraries, such as GAM, provide more efficient access to communication by removing the need to copy messages between the interface and user space in some cases. still lower-level interfaces, such as UNET, get right down to the hardware level to provide maximum performance. However, these are all still interfaces for passing messages from one process to another, and have limited utility in a shared-memory environment, due primarily to the fact that message passing is just another term for copying. This drawback is made more pertinent by today's hybrid architectures (e.g. clusters of SMPs), where it is difficult to know beforehand whether two communicating processes will share memory. As a result, even portable language tools (like HPF compilers) must either map all interprocess communication, into message passing with the accompanying performance degradation in shared memory environments, or they must check each communication at run-time and implement the shared-memory case separately for efficiency. Cooperative Data Sharing (CDS) is a single user-level API which abstracts all communication between processes into the sharing and access coordination of memory regions, in a model which might be described as "distributed shared messages" or "large-grain distributed shared memory". As a result, the user programs to a simple latency-tolerant abstract communication specification which can be mapped efficiently to either a shared-memory or message-passing based run-time system, depending upon the available architecture. Unlike some distributed shared memory interfaces, the user still has complete control over the assignment of data to processors, the forwarding of data to its next likely destination, and the queuing of data until it is needed, so even the relatively high latency present in clusters can be accomodated. CDS does not require special use of an MMU, which can add overhead to some DSM systems, and does not require an SPMD programming model. unlike some message-passing interfaces, CDS allows the user to implement efficient demand-driven applications where processes must "fight" over data, and does not perform copying if processes share memory and do not attempt concurrent writes. CDS also supports heterogeneous computing, dynamic process creation, handlers, and a very simple thread-arbitration mechanism. Additional support for array subsections is currently being considered. The CDS1 API, which forms the kernel of CDS, is built primarily upon only 2 communication primitives, one process initiation primitive, and some data translation (and marshalling) routines, memory allocation routines, and priority control routines. The entire current collection of 28 routines provides enough functionality to implement most (or all) of MPI 1 and 2, which has a much larger interface consisting of hundreds of routines. still, the API is small enough to consider integrating into standard os interfaces for handling inter-process communication in a network-independent way. This approach would also help to solve many of the problems plaguing other higher-level standards such as MPI and PVM which must, in some cases, "play OS" to adequately address progress and process control issues. The CDS2 API, a higher level of interface roughly equivalent in functionality to MPI and to be built entirely upon CDS1, is still being designed. It is intended to add support for the equivalent of communicators, reduction and other collective operations, process topologies, additional support for process creation, and some automatic memory management. CDS2 will not exactly match MPI, because the copy-free semantics of communication from CDS1 will be supported. CDS2 application programs will be free to carefully also use CDS1. CDS1 has been implemented on networks of workstations running unmodified Unix-based operating systems, using UDP/IP and vendor-supplied high- performance locks. Although its inter-node performance is currently unimpressive due to rudimentary implementation technique, it even now outperforms highly-optimized MPI implementation on intra-node communication due to its support for non-copy communication. The similarity of the CDS1 architecture to that of other projects such as UNET and TRAP suggests that the inter-node performance can be increased significantly to surpass MPI or PVM, and it may be possible to migrate some of its functionality to communication controllers.

DiNucci, David C.↗
DOE OSTI2023

A Sparse Distributed Gigascale Resolution Material Point Method

In this paper, we present a four-layer distributed simulation system and its adaptation to the Material Point Method (MPM). The system is built upon a performance portable C++ programming model targeting major High-Performance-Computing (HPC) platforms. A key ingredient of our system is a hierarchical block-tile-cell sparse grid data structure that is distributable to an arbitrary number of Message Passing Interface (MPI) ranks. We additionally propose strategies for efficient dynamic load balance optimization to maximize the efficiency of MPI tasks. Our simulation pipeline can easily switch among backend programming models, including OpenMP and CUDA, and can be effortlessly dispatched onto supercomputers and the cloud. Finally, we construct benchmark experiments and ablation studies on supercomputers and consumer workstations in a local network to evaluate the scalability and load balancing criteria. We demonstrate massively parallel, highly scalable, and gigascale resolution MPM simulations of up to 1.01 billion particles for less than 323.25 seconds per frame with 8 OpenSSH-connected workstations.

97 MATHEMATICS AND COMPUTING↗
NASA NTRS2000

Incremental Parallelization of Non-Data-Parallel Programs Using the Charon Message-Passing Library

Message passing is among the most popular techniques for parallelizing scientific programs on distributed-memory architectures. The reasons for its success are wide availability (MPI), efficiency, and full tuning control provided to the programmer. A major drawback, however, is that incremental parallelization, as offered by compiler directives, is not generally possible, because all data structures have to be changed throughout the program simultaneously. Charon remedies this situation through mappings between distributed and non-distributed data. It allows breaking up the parallelization into small steps, guaranteeing correctness at every stage. Several tools are available to help convert legacy codes into high-performance message-passing programs. They usually target data-parallel applications, whose loops carrying most of the work can be distributed among all processors without much dependency analysis. Others do a full dependency analysis and then convert the code virtually automatically. Even more toolkits are available that aid construction from scratch of message passing programs. None, however, allows piecemeal translation of codes with complex data dependencies (i.e. non-data-parallel programs) into message passing codes. The Charon library (available in both C and Fortran) provides incremental parallelization capabilities by linking legacy code arrays with distributed arrays. During the conversion process, non-distributed and distributed arrays exist side by side, and simple mapping functions allow the programmer to switch between the two in any location in the program. Charon also provides wrapper functions that leave the structure of the legacy code intact, but that allow execution on truly distributed data. Finally, the library provides a rich set of communication functions that support virtually all patterns of remote data demands in realistic structured grid scientific programs, including transposition, nearest-neighbor communication, pipelining, gather/scatter, and redistribution. At the end of the conversion process most intermediate Charon function calls will have been removed, the non-distributed arrays will have been deleted, and virtually the only remaining Charon functions calls are the high-level, highly optimized communications. Distribution of the data is under complete control of the programmer, although a wide range of useful distributions is easily available through predefined functions. A crucial aspect of the library is that it does not allocate space for distributed arrays, but accepts programmer-specified memory. This has two major consequences. First, codes parallelized using Charon do not suffer from encapsulation; user data is always directly accessible. This provides high efficiency, and also retains the possibility of using message passing directly for highly irregular communications. Second, non-distributed arrays can be interpreted as (trivial) distributions in the Charon sense, which allows them to be mapped to truly distributed arrays, and vice versa. This is the mechanism that enables incremental parallelization. In this paper we provide a brief introduction of the library and then focus on the actual steps in the parallelization process, using some representative examples from, among others, the NAS Parallel Benchmarks. We show how a complicated two-dimensional pipeline-the prototypical non-data-parallel algorithm- can be constructed with ease. To demonstrate the flexibility of the library, we give examples of the stepwise, efficient parallel implementation of nonlocal boundary conditions common in aircraft simulations, as well as the construction of the sequence of grids required for multigrid.

VanderWijngaart, Rob F.↗
NASA NTRS1999

Parallelizing OVERFLOW: Experiences, Lessons, Results

The computer code OVERFLOW is widely used in the aerodynamic community for the numerical solution of the Navier-Stokes equations. Current trends in computer systems and architectures are toward multiple processors and parallelism, including distributed memory. This report describes work that has been carried out by the author and others at Ames Research Center with the goal of parallelizing OVERFLOW using a variety of parallel architectures and parallelization strategies. This paper begins with a brief description of the OVERFLOW code. This description includes the basic numerical algorithm and some software engineering considerations. Next comes a description of a parallel version of OVERFLOW, OVERFLOW/PVM, using PVM (Parallel Virtual Machine). This parallel version of OVERFLOW uses the manager/worker style and is part of the standard OVERFLOW distribution. Then comes a description of a parallel version of OVERFLOW, OVERFLOW/MPI, using MPI (Message Passing Interface). This parallel version of OVERFLOW uses the SPMD (Single Program Multiple Data) style. Finally comes a discussion of alternatives to explicit message-passing in the context of parallelizing OVERFLOW.

Jespersen, Dennis C.↗
NASA NTRS2002

Effects of Ordering Strategies and Programming Paradigms on Sparse Matrix Computations

The Conjugate Gradient (CG) algorithm is perhaps the best-known iterative technique to solve sparse linear systems that are symmetric and positive definite. For systems that are ill-conditioned, it is often necessary to use a preconditioning technique. In this paper, we investigate the effects of various ordering and partitioning strategies on the performance of parallel CG and ILU(O) preconditioned CG (PCG) using different programming paradigms and architectures. Results show that for this class of applications: ordering significantly improves overall performance on both distributed and distributed shared-memory systems, that cache reuse may be more important than reducing communication, that it is possible to achieve message-passing performance using shared-memory constructs through careful data ordering and distribution, and that a hybrid MPI+OpenMP paradigm increases programming complexity with little performance gains. A implementation of CG on the Cray MTA does not require special ordering or partitioning to obtain high efficiency and scalability, giving it a distinct advantage for adaptive applications; however, it shows limited scalability for PCG due to a lack of thread level parallelism.

Oliker, Leonid↗