SEARCH · Engineering Papers

Results for “Message Passing”

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.

SourceResearch connectionFrom yearThrough year

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

At least 109 records · Page 6

All sources NASA NTRS DOE OSTI

NASA NTRS2002

Performance Evaluation of Remote Memory Access (RMA) Programming on Shared Memory Parallel Computers

The purpose of this study is to evaluate the feasibility of remote memory access (RMA) programming on shared memory parallel computers. We discuss different RMA based implementations of selected CFD application benchmark kernels and compare them to corresponding message passing based codes. For the message-passing implementation we use MPI point-to-point and global communication routines. For the RMA based approach we consider two different libraries supporting this programming model. One is a shared memory parallelization library (SMPlib) developed at NASA Ames, the other is the MPI-2 extensions to the MPI Standard. We give timing comparisons for the different implementation strategies and discuss the performance.

Jin, Hao-Qiang↗

DOE OSTI2011

Dispatching packets on a global combining network of a parallel computer

Methods, apparatus, and products are disclosed for dispatching packets on a global combining network of a parallel computer comprising a plurality of nodes connected for data communications using the network capable of performing collective operations and point to point operations that include: receiving, by an origin system messaging module on an origin node from an origin application messaging module on the origin node, a storage identifier and an operation identifier, the storage identifier specifying storage containing an application message for transmission to a target node, and the operation identifier specifying a message passing operation; packetizing, by the origin system messaging module, the application message into network packets for transmission to the target node, each network packet specifying the operation identifier and an operation type for the message passing operation specified by the operation identifier; and transmitting, by the origin system messaging module, the network packets to the target node.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2011

Global Arrays Parallel Programming Toolkit

The two predominant classes of programming models for parallel computing are distributed memory and shared memory. Both shared memory and distributed memory models have advantages and shortcomings. Shared memory model is much easier to use but it ignores data locality/placement. Given the hierarchical nature of the memory subsystems in modern computers this characteristic can have a negative impact on performance and scalability. Careful code restructuring to increase data reuse and replacing fine grain load/stores with block access to shared data can address the problem and yield performance for shared memory that is competitive with message-passing. However, this performance comes at the cost of compromising the ease of use that the shared memory model advertises. Distributed memory models, such as message-passing or one-sided communication, offer performance and scalability but they are difficult to program. The Global Arrays toolkit attempts to offer the best features of both models. It implements a shared-memory programming model in which data locality is managed by the programmer. This management is achieved by calls to functions that transfer data between a global address space (a distributed array) and local storage. In this respect, the GA model has similarities to the distributed shared-memory models that provide an explicit acquire/release protocol. However, the GA model acknowledges that remote data is slower to access than local data and allows data locality to be specified by the programmer and hence managed. GA is related to the global address space languages such as UPC, Titanium, and, to a lesser extent, Co-Array Fortran. In addition, by providing a set of data-parallel operations, GA is also related to data-parallel languages such as HPF, ZPL, and Data Parallel C. However, the Global Array programming model is implemented as a library that works with most languages used for technical computing and does not rely on compiler technology for achieving parallel efficiency. It also supports a combination of task- and data-parallelism and is available as an extension of the message passing (MPI) model. The GA model exposes to the programmer the hierarchical memory of modern high-performance computer systems, and by recognizing the communication overhead for remote data transfer, it promotes data reuse and locality of reference. Virtually all the scalable architectures possess non-uniform memory access characteristics that reflect their multi-level memory hierarchies. These hierarchies typically comprise processor registers, multiple levels of cache, local memory, and remote memory. Over time, both the number of levels and the cost (in processor cycles) of accessing deeper levels has been increasing. It is important for any scalable programming model to address memory hierarchy since it is critical to the efficient execution of scalable applications.

ARMCI↗

NASA NTRS1996

Parallel NPARC: Implementation and Performance

Version 3 of the NPARC Navier-Stokes code includes support for large-grain (block level) parallelism using explicit message passing between a heterogeneous collection of computers. This capability has the potential for significant performance gains, depending upon the block data distribution. The parallel implementation uses a master/worker arrangement of processes. The master process assigns blocks to workers, controls worker actions, and provides remote file access for the workers. The processes communicate via explicit message passing using an interface library which provides portability to a number of message passing libraries, such as PVM (Parallel Virtual Machine). A Bourne shell script is used to simplify the task of selecting hosts, starting processes, retrieving remote files, and terminating a computation. This script also provides a simple form of fault tolerance. An analysis of the computational performance of NPARC is presented, using data sets from an F/A-18 inlet study and a Rocket Based Combined Cycle Engine analysis. Parallel speedup and overall computational efficiency were obtained for various NPARC run parameters on a cluster of IBM RS6000 workstations. The data show that although NPARC performance compares favorably with the estimated potential parallelism, typical data sets used with previous versions of NPARC will often need to be reblocked for optimum parallel performance. In one of the cases studied, reblocking increased peak parallel speedup from 3.2 to 11.8.

Townsend, S. E.↗

NASA NTRS1999

Performance and Application of Parallel OVERFLOW Codes on Distributed and Shared Memory Platforms

The presentation discusses recent studies on the performance of the two parallel versions of the aerodynamics CFD code, OVERFLOW_MPI and _MLP. Developed at NASA Ames, the serial version, OVERFLOW, is a multidimensional Navier-Stokes flow solver based on overset (Chimera) grid technology. The code has recently been parallelized in two ways. One is based on the explicit message-passing interface (MPI) across processors and uses the _MPI communication package. This approach is primarily suited for distributed memory systems and workstation clusters. The second, termed the multi-level parallel (MLP) method, is simple and uses shared memory for all communications. The _MLP code is suitable on distributed-shared memory systems. For both methods, the message passing takes place across the processors or processes at the advancement of each time step. This procedure is, in effect, the Chimera boundary conditions update, which is done in an explicit "Jacobi" style. In contrast, the update in the serial code is done in more of the "Gauss-Sidel" fashion. The programming efforts for the _MPI code is more complicated than for the _MLP code; the former requires modification of the outer and some inner shells of the serial code, whereas the latter focuses only on the outer shell of the code. The _MPI version offers a great deal of flexibility in distributing grid zones across a specified number of processors in order to achieve load balancing. The approach is capable of partitioning zones across multiple processors or sending each zone and/or cluster of several zones into a single processor. The message passing across the processors consists of Chimera boundary and/or an overlap of "halo" boundary points for each partitioned zone. The MLP version is a new coarse-grain parallel concept at the zonal and intra-zonal levels. A grouping strategy is used to distribute zones into several groups forming sub-processes which will run in parallel. The total volume of grid points in each group are approximately balanced. A proper number of threads are initially allocated to each group, and in subsequent iterations during the run-time, the number of threads are adjusted to achieve load balancing across the processes. Each process exploits the multitasking directives already established in Overflow.

Djomehri, M. Jahed↗

NASA NTRS1994

High Performance Programming Using Explicit Shared Memory Model on Cray T3D1

The Cray T3D system is the first-phase system in Cray Research, Inc.'s (CRI) three-phase massively parallel processing (MPP) program. This system features a heterogeneous architecture that closely couples DEC's Alpha microprocessors and CRI's parallel-vector technology, i.e., the Cray Y-MP and Cray C90. An overview of the Cray T3D hardware and available programming models is presented. Under Cray Research adaptive Fortran (CRAFT) model four programming methods (data parallel, work sharing, message-passing using PVM, and explicit shared memory model) are available to the users. However, at this time data parallel and work sharing programming models are not available to the user community. The differences between standard PVM and CRI's PVM are highlighted with performance measurements such as latencies and communication bandwidths. We have found that the performance of neither standard PVM nor CRI s PVM exploits the hardware capabilities of the T3D. The reasons for the bad performance of PVM as a native message-passing library are presented. This is illustrated by the performance of NAS Parallel Benchmarks (NPB) programmed in explicit shared memory model on Cray T3D. In general, the performance of standard PVM is about 4 to 5 times less than obtained by using explicit shared memory model. This degradation in performance is also seen on CM-5 where the performance of applications using native message-passing library CMMD on CM-5 is also about 4 to 5 times less than using data parallel methods. The issues involved (such as barriers, synchronization, invalidating data cache, aligning data cache etc.) while programming in explicit shared memory model are discussed. Comparative performance of NPB using explicit shared memory programming model on the Cray T3D and other highly parallel systems such as the TMC CM-5, Intel Paragon, Cray C90, IBM-SP1, etc. is presented.

Simon, Horst D.↗

NASA NTRS1994

Performance Measurement, Visualization and Modeling of Parallel and Distributed Programs

This paper presents a methodology for debugging the performance of message-passing programs on both tightly coupled and loosely coupled distributed-memory machines. The AIMS (Automated Instrumentation and Monitoring System) toolkit, a suite of software tools for measurement and analysis of performance, is introduced and its application illustrated using several benchmark programs drawn from the field of computational fluid dynamics. AIMS includes (i) Xinstrument, a powerful source-code instrumentor, which supports both Fortran77 and C as well as a number of different message-passing libraries including Intel's NX Thinking Machines' CMMD, and PVM; (ii) Monitor, a library of timestamping and trace -collection routines that run on supercomputers (such as Intel's iPSC/860, Delta, and Paragon and Thinking Machines' CM5) as well as on networks of workstations (including Convex Cluster and SparcStations connected by a LAN); (iii) Visualization Kernel, a trace-animation facility that supports source-code clickback, simultaneous visualization of computation and communication patterns, as well as analysis of data movements; (iv) Statistics Kernel, an advanced profiling facility, that associates a variety of performance data with various syntactic components of a parallel program; (v) Index Kernel, a diagnostic tool that helps pinpoint performance bottlenecks through the use of abstract indices; (vi) Modeling Kernel, a facility for automated modeling of message-passing programs that supports both simulation -based and analytical approaches to performance prediction and scalability analysis; (vii) Intrusion Compensator, a utility for recovering true performance from observed performance by removing the overheads of monitoring and their effects on the communication pattern of the program; and (viii) Compatibility Tools, that convert AIMS-generated traces into formats used by other performance-visualization tools, such as ParaGraph, Pablo, and certain AVS/Explorer modules.

Yan, Jerry C.↗

DOE OSTI2016

Multi-node and Multi-core Performance Studies of a Monte Carlo code RMC

Over the past 15 years the architecture of high-performance computing systems has become increasingly heterogeneous with individual computing cores being subdivided into nodes and even auxiliary co-processing units such as graphics processing units or the Intel many integrated core co-processors. In fact, future computer platforms is moving toward a larger number of nodes and cores with lower memory available per node. Today, several production Monte Carlo codes use some types of hybrid parallelism strategy that makes use of shared-memory parallelism within a computing node and message-passing parallelism between nodes, which enables the codes to utilize available computing resources as much as possible, both among and within computing nodes on a cluster. The so-called hybrid parallelization is also possible with Reactor Monte Carlo Code RMC by activating both MPI and OpenMP parallel algorithms together. On HPC' which typically consists of several or several-ten thousand nodes equipping 16-32 computing intra-cores per node, simulation of RMC can be carried out in the hybrid parallelization with the MPI inter-node parallelization and the OpenMP intra-node parallelization. RMC has adopted a hybrid message-passing and shared-memory approach to parallelism, which enables the code to utilize available computing resources, both among and within computing nodes on a cluster. Shared-memory parallelism is especially valuable for large-model simulations because it enables multiple compute cores within a node to share common model information such as the geometry description, material, cross section and tally data, thus reducing the memory usage per processor. The calculation on the Tianhe-2 supercomputer took 6144 cores, resulting in a rate of 173,314,000 neutrons/minute for BEAVRS benchmark. While pin-wise full core H-M benchmark burnup calculations with millions of depletion regions was selected using MPI/OpenMP parallelism, the speed of 6,144 cores reach 1,843,200 burnable regions per minute. Keeping up with the continuing trend for increases in the number of computing resources within modern HPC systems is a significant challenge for teams that develop and maintain high-performance MC radiation transport solvers. order to address this challenge, RMC has adopted a hybrid message-passing and shared-memory approach to parallelism, which enables the code to utilize available computing resources, both among and within computing nodes on a cluster. Shared-memory parallelism is especially valuable for large-model simulations because it enables multiple compute cores within a node to share common model information such as the geometry description, material and composition definitions, and cross section data, thus reducing the memory usage per processor. The parallel performance has been studied by using two complex and representative benchmarks, H-M and BEAVRS benchmark. All simulations were run on Intel Xeon E5-2692 12 core, 2.2 GHz processors at MilkyWay-2 Super Computer with up to 12,288 available computing cores. While employing MPI parallelism, excellent parallel efficiency is achieved even above 12,000 processors and RMC reaches the parallel efficiency of 96% on 12,288 cores for the pin-by-pin computation of the BEAVRS benchmark. While BEAVRS criticality benchmark was selected using MPI/OpenMP parallelism, the speed of 6,144 cores reach 173,314,000 particles per minute. While pin-wise full core H-M benchmark burnup calculations with millions of depletion regions was selected using MPI/OpenMP parallelism, the speed of 6,144 cores reach 1,843,200 burnable regions per minute. Results from these studies demonstrate that RMC, using the hybrid parallel model described in this paper, is able to handle pin-wise full core burnup calculations with millions of depletion regions and scales well through thousands of processors. (authors)

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗

NASA NTRS1998

Navier-Stokes Computations on Commodity Computers

In this paper we discuss and demonstrate the feasibility of solving high-fidelity, nonlinear computational fluid dynamics (CFD) problems of practical interest on commodity machines, namely Pentium Pro PC's. Such calculations have now become possible due to the progress in computational power and memory of the off-the-shelf commodity computers, along with the growth in bandwidth and communication speeds of networks. A widely used CFD code known as TLNS3D, which was developed originally on large shared memory computers was selected for this effort. This code has recently been ported to massively parallel processor (MPP) type machines, where natural partitioning along grid blocks is adopted in which one or more blocks are distributed to each of the available processors. In this paper, a similar approach is adapted to port this code to a cluster of Pentium Pro computers. The message passing among the processors is accomplished through the use of standard message passing interface (MPI) libraries. Scaling studies indicate fairly high level of parallelism on such clusters of commodity machines, thus making solutions to Navier-Stokes equations for practical problems more affordable.

Vatsa, Veer N.↗

DOE OSTI2026

When does global attention help: a unified empirical study on atomistic graph learning

Graph neural networks (GNNs) are widely used as surrogates for costly experiments and first-principles simulations to study the behavior of compounds at atomistic scale, and their architectural complexity is constantly increasing to enable the modeling of complex physics. While most recent GNNs combine more traditional message passing neural networks (MPNNs) layers to model short-range interactions with more advanced graph transformers (GTs) with global attention mechanisms to model long-range interactions, it is still unclear when global attention mechanisms provide real benefits over well-tuned MPNN layers due to inconsistent implementations, features, or hyperparameter tuning. We introduce the first unified, reproducible benchmarking framework–built on HydraGNN–that enables seamless switching among four controlled model classes: MPNN, MPNN with chemistry/topology encoders, GPS-style hybrids of MPNN with global attention, and fully fused localglobal models with encoders. Using seven diverse open-source datasets for benchmarking across regression and classification tasks, we systematically isolate the contributions of message passing, global attention, and encoder-based feature augmentation. Our study shows that encoder-augmented MPNNs form a robust baseline, while fused localglobal models yield the clearest benefits for properties governed by long-range interaction effects. We further quantify the accuracycompute trade-offs of attention, reporting its overhead in memory. Together, these results establish the first controlled evaluation of global attention in atomistic graph learning and provide a reproducible testbed for future model development.

Equivariant graph neural networks↗

DOE OSTI2012

Compiled MPI: Cost-Effective Exascale Applications Development

The complexity of petascale and exascale machines makes it increasingly difficult to develop applications that can take advantage of them. Future systems are expected to feature billion-way parallelism, complex heterogeneous compute nodes and poor availability of memory (Peter Kogge, 2008). This new challenge for application development is motivating a significant amount of research and development on new programming models and runtime systems designed to simplify large-scale application development. Unfortunately, DoE has significant multi-decadal investment in a large family of mission-critical scientific applications. Scaling these applications to exascale machines will require a significant investment that will dwarf the costs of hardware procurement. A key reason for the difficulty in transitioning today's applications to exascale hardware is their reliance on explicit programming techniques, such as the Message Passing Interface (MPI) programming model to enable parallelism. MPI provides a portable and high performance message-passing system that enables scalable performance on a wide variety of platforms. However, it also forces developers to lock the details of parallelization together with application logic, making it very difficult to adapt the application to significant changes in the underlying system. Further, MPI's explicit interface makes it difficult to separate the application's synchronization and communication structure, reducing the amount of support that can be provided by compiler and run-time tools. This is in contrast to the recent research on more implicit parallel programming models such as Chapel, OpenMP and OpenCL, which promise to provide significantly more flexibility at the cost of reimplementing significant portions of the application. We are developing CoMPI, a novel compiler-driven approach to enable existing MPI applications to scale to exascale systems with minimal modifications that can be made incrementally over the application's lifetime. It includes: (1) New set of source code annotations, inserted either manually or automatically, that will clarify the application's use of MPI to the compiler infrastructure, enabling greater accuracy where needed; (2) A compiler transformation framework that leverages these annotations to transform the original MPI source code to improve its performance and scalability; (3) Novel MPI runtime implementation techniques that will provide a rich set of functionality extensions to be used by applications that have been transformed by our compiler; and (4) A novel compiler analysis that leverages simple user annotations to automatically extract the application's communication structure and synthesize most complex code annotations.

97 MATHEMATICS AND COMPUTING↗

NASA NTRS2008

PyPele Rewritten To Use MPI

A computer program known as PyPele, originally written as a Pythonlanguage extension module of a C++ language program, has been rewritten in pure Python language. The original version of PyPele dispatches and coordinates parallel-processing tasks on cluster computers and provides a conceptual framework for spacecraft-mission- design and -analysis software tools to run in an embarrassingly parallel mode. The original version of PyPele uses SSH (Secure Shell a set of standards and an associated network protocol for establishing a secure channel between a local and a remote computer) to coordinate parallel processing. Instead of SSH, the present Python version of PyPele uses Message Passing Interface (MPI) [an unofficial de-facto standard language-independent application programming interface for message- passing on a parallel computer] while keeping the same user interface. The use of MPI instead of SSH and the preservation of the original PyPele user interface make it possible for parallel application programs written previously for the original version of PyPele to run on MPI-based cluster computers. As a result, engineers using the previously written application programs can take advantage of embarrassing parallelism without need to rewrite those programs.

Hockney, George↗

DOE OSTI2020

Machine Learning Methods for Connection RTT and Loss Rate Estimation Using MPI Measurements Under Random Losses

Scientific computations are expected to be increasingly distributed across wide-area networks, and Message Passing Interface (MPI) has been shown to scale to support their communications over long distances. Application-level measurements of MPI operations reflect the connection Round-Trip Time (RTT) and loss rate, and machine learning methods have been previously developed to estimate them under deterministic periodic losses. In this paper, we consider more complex, random losses with unform, Poisson and Gaussian distributions. We study five disparate machine leaning methods, with linear and non-linear, and smooth and non-smooth properties, to estimate RTT and loss rate over 10Gbps connections with 0-366ms RTT. The diversity and complexity of these estimators combined with the randomness of losses and TCP’s non-linear response together rule out the selection of a single best among them; instead, we fuse them to retain their design diversity. Overall, the results show that accurate estimates can be generated at low loss rates but become inaccurate at loss rates 10% and higher, thereby illustrating both their strengths and limitations.

Generalization Bounds↗

NASA NTRS2003

Comparing the OpenMP, MPI, and Hybrid Programming Paradigm on an SMP Cluster

Clusters of SMP (Symmetric Multi-Processors) nodes provide support for a wide range of parallel programming paradigms. The shared address space within each node is suitable for OpenMP parallelization. Message passing can be employed within and across the nodes of a cluster. Multiple levels of parallelism can be achieved by combining message passing and OpenMP parallelization. Which programming paradigm is the best will depend on the nature of the given problem, the hardware components of the cluster, the network, and the available software. In this study we compare the performance of different implementations of the same CFD benchmark application, using the same numerical algorithm but employing different programming paradigms.

Jost, Gabriele↗

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↗

DOE OSTI2023

Multiscale graph neural network autoencoders for interpretable scientific machine learning

The goal of this work is to address two limitations in autoencoder-based models: latent space interpretability and compatibility with unstructured meshes. This is accomplished here with the development of a novel graph neural network (GNN) autoencoding architecture with demonstrations on complex fluid flow applications. To address the first goal of interpretability, the GNN autoencoder achieves reduction in the number nodes in the encoding stage through an adaptive graph reduction procedure. Further, this reduction procedure essentially amounts to flowfieldconditioned node sampling and sensor identification, and produces interpretable latent graph representations tailored to the flowfield reconstruction task in the form of so-called masked fields. These masked fields allow the user to (a) visualize where in physical space a given latent graph is active, and (b) interpret the time-evolution of the latent graph connectivity in accordance with the time-evolution of unsteady flow features (e.g. recirculation zones, shear layers) in the domain. To address the goal of unstructured mesh compatibility, the autoencoding architecture utilizes a series of multi-scale message passing (MMP) layers, each of which models information exchange among node neighborhoods at various lengthscales. The MMP layer, which augments standard single-scale message passing with learnable coarsening operations, allows the decoder to more efficiently reconstruct the flowfield from the identified regions in the masked fields. Analysis of latent graphs produced by the autoencoder for various model settings are conducted using unstructured snapshot data sourced from large-eddy simulations in a backward-facing step (BFS) flow configuration with an OpenFOAM-based flow solver at high Reynolds numbers.

97 MATHEMATICS AND COMPUTING↗

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 NTRS1993

Evaluation Metrics for the Paragon XP/S-15

On February 17th 1993, the Numerical Aerodynamic Simulation (NAS) facility located at the NASA Ames Research Center installed a 224 node Intel Paragon XP/S-15 system. After its installation, the Paragon was found to be in a very immature state and was unable to support a NAS users' workload, composed of a wide range of development and production activities. As a first step towards addressing this problem, we implemented a set of metrics to objectively monitor the system as operating system and hardware upgrades were installed. The metrics were designed to measure four aspects of the system that we consider essential to support our workload: availability, utilization, functionality, and performance. This report presents the metrics collected from February 1993 to August 1993. Since its installation, the Paragon availability has improved from a low of 15% uptime to a high of 80%, while its utilization has remained low. Functionality and performance have improved from merely running one of the NAS Parallel Benchmarks to running all of them faster (between 1 and 2 times) than on the iPSC/860. In spite of the progress accomplished, fundamental limitations of the Paragon operating system are restricting the Paragon from supporting the NAS workload. The maximum operating system message passing (NORMA IPC) bandwidth was measured at 11 Mbytes/s, well below the peak hardware bandwidth (175 Mbytes/s), limiting overall virtual memory and Unix services (i.e. Disk and HiPPI I/O) performance. The high NX application message passing latency (184 microns), three times than on the iPSC/860, was found to significantly degrade performance of applications relying on small message sizes. The amount of memory available for an application was found to be approximately 10 Mbytes per node, indicating that the OS is taking more space than anticipated (6 Mbytes per node).

Traversat, Bernard↗