Large-scale Parallel Computers Research Articles

An algebraic approach for data-centric scientific workflows

Scientific workflows have emerged as a basic abstraction for structuring and executing scientific experiments in computational environments. In many situations, these workflows are computationally and data intensive, thus requiring execution in large-scale parallel computers. However, parallelization of scientific workflows remains low-level, ad-hoc and labor-intensive, which makes it hard to exploit optimization opportunities. To address this problem, we propose an algebraic approach (inspired by relational algebra) and a parallel execution model that enable automatic optimization of scientific workflows. We conducted a thorough validation of our approach using both a real oil exploitation application and synthetic data scenarios. The experiments were run in Chiron, a data-centric scientific workflow engine implemented to support our algebraic approach. Our experiments demonstrate performance improvements of up to 226% compared to an ad-hoc workflow implementation.

Dynamic Fault Tolerance in Fat Trees

Fat trees are a very common communication architecture in current large-scale parallel computers. The probability of failure in these systems increases with the number of components. We present a routing method for deterministically and adaptively routed fat trees, applicable to both distributed and source routing, that is able to handle several concurrent faults and that transparently returns to the original routing strategy once the faulty components have recovered. The method is local and dynamic, completely masking the fault from the rest of the system. It only requires a small extra functionality in the switches to handle rerouting packets around a fault. The method guarantees connectedness and deadlock and livelock freedom for up to k -1 benign simultaneous switch and/or link faults where k is half the number of ports in the switches. Our simulation experiments show a graceful degradation of performance as more faults occur. Furthermore, we demonstrate that for most fault combinations, our method will even be able to handle significantly more faults beyond the k -1 limit with high probability.

Framework for Massively Parallel Adaptive Finite Element Computational Fluid Dynamics on Tetrahedral Meshes

In this paper we describe a general adaptive finite element framework for unstructured tetrahedral meshes without hanging nodes suitable for large scale parallel computations. Our framework is designed to scale linearly to several thousands of processors, using fully distributed and efficient algorithms. The key components of our implementation, local mesh refinement and load balancing algorithms, are described in detail. Finally, we present a theoretical and experimental performance study of our framework, used in a large scale computational fluid dynamics computation, and we compare scaling and complexity of different algorithms on different massively parallel architectures.

1512 多数の高分子モデルが分散した一様乱流の統計性(OS15.固体物理/流体物理のマルチフィジックス/マルチスケール解析(3),OS・一般セッション講演)

We examined the effects of the polymer additives on statistically steady and homogeneous turbulence by direct numerical simulations (DNS) of turbulence coupled with the Brownian dynamics (BD) simulations of the enormous number of FENE-dumbbells. We observed a reduction of the energy dissipation rate and a decrease of the acceleration variance of fluid particle by the effects of polymer additives. We found that these were consistent with the results of recent experimental and numerical studies. The efficiency of the large-scale parallel computations was also examined.

Protection Engineering Construction and Capability Analysis of FDS Parallel Computing Environment

The fire field model FDS can get more details of distribution of physical quantities and its process, so FDS has been widely used in the Performance-based design. To improve computing speed of the software, Based on MPI (Message Passing Interface) using the VC++language, FDS parallel computing program has been compiled. Specific to The fluid field using multiregional dissected and each sub segment would be allocated appropriate panel point to calculate. The calculation shows that with the increased amount of processor computing time has significantly reduced. The procedure and method that the text has established was feasible, and could further extend to large-scale parallel computing and engineering application.

Linux Cluster in Theory and Practice: A Novel Approach in Teaching Cluster Computing Based on the Intel Atom Platform

Abstract Current trends and studies on future architectures show, that the complexity of parallel computer systems is increasing steadily. Hence, the industry requires skilled employees, who have in addition to the theoretical fundamentals, practical experiences in the design and administration of such systems. However, investigations have shown, that practical approaches are still missing in current curricula, especially in these areas. For this reason, the chair of Computer Architecture at the faculty of Computer Science at the Technische Universiẗat Dresden, developed and introduced the course “Linux Cluster in Theory and Practice” (LCTP). The main objectives of this course are to provide background knowledge about the design and administration of large-scale parallel computer systems and the practical implementation on the available hardware. In addition, students learn how to solve problems in a structured approach and as part of a team. This paper analyzes the current variety of courses in the area of parallel computing systems, describes the structure and implementation of LCTP and provides first conclusions and an outlook on possible further developments.

A High Performance/Price Ratio Cluster Computing Platform for Simulation Design

Based on our Fibre-Channel Token-Routing Switch-Network, a high performance/price ratio cluster computing platform is put forward and developed for simulation design which needs large-scale parallel computation. According to the feature of our cluster computing platform, the cluster computing software platform is proposed and implemented. The cluster computing software platform has two parts: the cluster management scheduling sub-platform and the cluster tasks running sub-platform. The cluster management scheduling sub-platform is composed of the cluster management system and the task scheduling system, and the cluster tasks running sub-platform includes the FC-TR communication software and the FC-TR parallel programming environment. The FC-TR communication software provides high-bandwidth, low-latency communication services for the parallel programming environment on the upper layer, while the parallel programming environment makes a number of large-scale-computation simulation design applications run on our network seamlessly. The experiment results show that our cluster computing platform has achieved better performance than the Scali cluster computing platform.

Efficient Numerical Approach to Inhomogeneous Superconductivity: The Chebyshev-Bogoliubov–de Gennes Method

We propose a highly efficient numerical method to describe inhomogeneous superconductivity by using the kernel polynomial method in order to calculate the Green's functions of a superconductor. Broken translational invariance of any type (impurities, surfaces, or magnetic fields) can be easily incorporated. We show that limitations due to system size can be easily circumvented and therefore this method opens the way for the study of scenarios and/or geometries that were unaccessible before. The proposed method is highly efficient and amenable to large scale parallel computation. Although we only use it in the context of superconductivity, it is applicable to other inhomogeneous mean-field theories.

Fine-grained parallel RNA secondary structure prediction using SCFGs on FPGA

In the field of RNA secondary structure prediction, the CYK (Coche–Younger–Kasami) algorithm is one of the most popular methods using a SCFG (stochastic context-free grammar) model. Accelerating SCFGs for large models and large RNA database searching becomes a challenging task in computational bioinformatics because the parallel efficiency of general purpose computer systems is limited by the O ( L 3) computational complexity and by complicated data dependences. Furthermore, large scale parallel computers are too expensive to be easily accessible to many research institutes. Recently, FPGA chips have emerged as one promising application accelerator to accelerate the CYK algorithm by exploiting a fine-grained custom design. We propose a systolic-like array structure including one master PE and multiple slave PEs for the fine-grained hardware implementation on FPGA to accelerate the CYK/inside algorithm with Query-Dependent Banding (QDB) heuristics. We partition the tasks by columns and assign them to PEs for load balance. We exploit data reuse schemes to reduce the need to load matrices from external memory. The experimental results show a speedup factor of more than 14x over the Infernal−1.0 with QDB optimization for the alignment of a single long RNA sequence to a large CM model with thousands of states running on a PC platform with Intel Dual-core 2.5 GHz CPU. The computational power of our accelerator is comparable to that of a PC cluster consisting of 16 Intel-Xeon 2.0 GHz Quad CPUs for large-scale database alignment applications ( cmsearch) with multiple input sequences, but the power consumption is only about 10% of that of the cluster.

PHANTOM

For designers of large-scale parallel computers, it is greatly desired that performance of parallel applications can be predicted at the design phase. However, this is difficult because the execution time of parallel applications is determined by several factors, including sequential computation time in each process, communication time and their convolution. Despite previous efforts, it remains an open problem to estimate sequential computation time in each process accurately and efficiently for large-scale parallel applications on non-existing target machines. This paper proposes a novel approach to predict the sequential computation time accurately and efficiently. We assume that there is at least one node of the target platform but the whole target system need not be available. We make two main technical contributions. First, we employ deterministic replay techniques to execute any process of a parallel application on a single node at real speed. As a result, we can simply measure the real sequential computation time on a target node for each process one by one. Second, we observe that computation behavior of processes in parallel applications can be clustered into a few groups while processes in each group have similar computation behavior. This observation helps us reduce measurement time significantly because we only need to execute representative parallel processes instead of all of them. We have implemented a performance prediction framework, called PHANTOM, which integrates the above computation-time acquisition approach with a trace-driven network simulator. We validate our approach on several platforms. For ASCI Sweep3D, the error of our approach is less than 5% on 1024 processor cores. Compared to a recent regression-based prediction approach, PHANTOM presents better prediction accuracy across different platforms.

An effective speedup metric for measuring productivity in large-scale parallel computer systems

With the parallel computer systems scaling-up, the measure index for performance of the systems demands a shift from traditional "high performance" to "high productivity." This brings a new challenge to defining a synthetic, yet meaningful, measure index of multiple productivity variables; namely computing performance, reliability, energy consumption, parallel software development, etc. Traditional measures for large-scale parallel computer systems merely focus on computing performance, and are incapable of measuring the multiple productivity variables simultaneously in an effective manner. A recently proposed market-related money model, which pursues high utility/cost ratio, relies on money as a measure to consider the multiple productivity variables. Differing from the previous models, this paper proposes a novel system productivity speedup metric for large-scale parallel computer systems. The metric uses speedup instead of money to comprehensively unify the measures of multiple productivity variables. Finally, we propose a trade-off productivity measurement to weigh different productivity variables, to address different design targets. The measurement can facilitate the system evaluation, expose future technique tendencies, and guide future system design.

Lattice Boltzmann simulation on Nvidia and AMD GPUs

General purpose computing on graphic processing units (GPUs) has received great attention re-cently. Both Nvidia and AMD have announced their own SDK, CUDA and ASC, respectively. Com-pared with CPUs, many applications can achieve high speed-up on GPUs. As a mesh-based particle method for flow simulation, lattice Boltzmann method (LBM) has perfect intrinsic parallelism that is very suitable for large-scale parallel computing. In this article, we put forward an algorithm for LB simulation of flow in grooved channel using the D2Q9 model, running on both Nvidia and AMD GPUs in the multi-program multi-data (MPMD) mode through message passing interface (MPI). The capability of hybrid GPU clusters can thus be fully utilized. The correctness and performance of the computation is analyzed through comparison with corresponding CPU implementation and significant speed-up of the GPU implementation has been demonstrated. The results provide valuable references to the establishment of GPU-based high performance computing (HPC) systems.

A parallel algorithm with interface prediction and correction for spherical geometric transport equation

When solving transport equation by the discrete ordinates method, the solution procedure must accord with the moving direction of particle due to the stability restriction. Thus it results in a strong correlation among these unknown variables, and the intrinsic serial nature of the corresponding algorithmic procedure becomes the bottleneck for transport equation calculations on large-scale parallel computers. The conventional parallel sweep method has restrictive but in no means high degree of parallelism for two and three spatial dimensional problems, and it has no spatial parallel degree for one-dimensional problems. In this paper, a spatial domain decomposition method is adopted in the computational domain, and different interface prediction and correction methods are introduced to solve the 1-D spherical geometric transport equation. It is natural to generalize to multidimension problems. In order to avoid producing negative flux, we employ the exponential method for variable discretization. The numerical experiments by using MPI show that our parallel method with explicit prediction and implicit correction has good precision, parallelism and simplicity.

Outflow Boundary Conditions for Arterial Networks with Multiple Outlets

Simulation of blood flow in three-dimensional geometrically complex arterial networks involves many inlets and outlets and requires large-scale parallel computing. It should be based on physiologically correct boundary conditions, which are accurate, robust, and simple to implement in the parallel framework. While a secondary closure problem can be solved to provide approximate outflow conditions, it is preferable, when possible, to impose the clinically measured flow rates. We have developed a new method to incorporate such measurements at multiple outlets, based on a time-dependent resistance boundary condition for the pressure in conjunction with a Neumann boundary condition for the velocity. Convergence of the numerical solution for the specified outlet flow rates is achieved very fast at a computational complexity comparable to the widely used Resistance or Windkessel boundary conditions. The method is verified using a patient-specific cranial vascular network involving 20 arteries and 10 outlets.

Large-Scale Parallel Computation Methodologies for Highly Nonlinear Concrete and Soil Applications

Detailed analyses of concrete and buried concrete structures undergoing complex inelastic responses to loads, such as those resulting from explosive detonations, are challenging mechanics problems and can require significant computational resources. The writers have been involved in the development of various constitutive models that are successful in modeling blast responses, but can also be computationally intensive—thus excluding their use for many large-scale applications. Recent efforts at the U.S. Army Engineer Research and Development Center have focused on developing procedures for performing these types of analyses in a production setting utilizing high performance computing. These models have been implemented into a parallel finite-element code, ParaAble, developed by the writers, and a new feature was added to the METIS partitioning software to easily apply weighting for improved load balancing in multiple material problems. Examples are shown that efficiently utilize from dozens to up to thous...

Assembling genomes on large-scale parallel computers

Assembly of large genomes from tens of millions of short genomic fragments is computationally demanding requiring hundreds of gigabytes of memory and tens of thousands of CPU hours. The advent of high throughput sequencing technologies, new gene-enrichment sequencing strategies, and collective sequencing of environmental samples further exacerbate this situation. In this paper, we present the first massively parallel genome assembly framework. The unique features of our approach include space-efficient and on-demand algorithms that consume only linear space, and strategies to reduce the number of expensive pairwise sequence alignments while maintaining assembly quality. Developed as part of the ongoing efforts in maize genome sequencing, we applied our assembly framework to genomic data containing a mixture of gene enriched and random shotgun sequences. We report the partitioning of more than 1.6 million fragments of over 1.25 billion nucleotides total size into genomic islands in under 2 h on 1024 processors of an IBM BlueGene/L supercomputer. We also demonstrate the effectiveness of the proposed approach for traditional whole genome shotgun sequencing and assembly of environmental sequences.

Performance analysis of fault-tolerant routing algorithm in wormhole-switched interconnections

With nowadays popularity of large-scale parallel computers, Multiprocessors System-on-Chip (MP-SoCs), multicomputers, cluster computers and peer-to-peer communication networks, fault-tolerant routing becomes an important issue in developing these systems. Fault-tolerant routing algorithms in such systems aim at providing continuous operations in the presence of one or more failures by allowing the graceful degradation of system. The Software-Based fault-tolerant routing scheme has been suggested as an efficient routing algorithm to preserve both communication performance and fault-tolerant demands in parallel computer systems. To study network performance, a number of different analytical models for fault-free routing algorithms have been proposed in the past literature. However, there has not been reported any similar analytical model of fault-tolerant routing in the presence of faulty components. This paper presents a new analytical modeling approach for determining the effects of failures in wormhole-switched 2-D tori using the fault-tolerant Software-Based scheme. More specifically, we describe a general model to derive mathematical expressions to investigate the performance behavior of routing algorithms confronting convex (|-shaped, ?-shaped) or concave (U-shaped, +-shaped, T-shaped, H-shaped) faulty regions. The model is validated through comprehensive simulation experiments for different types of failures.

Parallel Computation of Large-Scale Molecular Dynamics Simulations

A large-scale parallel computation is extremely important for MD (molecular dynamics) simulation, particularly in dealing with atomistic systems with realistic size comparable to macroscopic continuum scale. We present a new approach for parallel computation of MD simulation. The entire system domain under consideration is divided into many Eulerian subdomains, each of which is surrounded with its own buffer layer and to which its own processor is assigned. This leads to an efficient tracking of each molecule, even when the molecules move out of its subdomain. Several numerical examples are provided to demonstrate the effectiveness of this computation scheme.

Parallel clustering algorithms for structured AMR

We compare several different parallel implementation approaches for the clustering operations performed during adaptive meshing operations in patch-based structured adaptive mesh refinement (SAMR) applications. Specifically, we target the clustering algorithm of Berger and Rigoutsos, which is commonly used in many SAMR applications. The baseline for comparison is a single program, multiple data extension of the original algorithm that works well for up to O ( 10 2 ) processors. Our goal is a clustering algorithm for machines of up to O ( 10 5 ) processors, such as the 64K-processor IBM BlueGene/L (BG/L) system. We first present an algorithm that avoids unneeded communications of the baseline approach, improving the clustering speed by up to an order of magnitude. We then present a new task-parallel implementation to further reduce communication wait time, adding another order of magnitude of improvement. The new algorithms exhibit more favorable scaling behavior for our test problems. Performance is evaluated on a number of large-scale parallel computer systems, including a 16K-processor BG/L system.

Numerical simulation of a mini PEMFC stack

Fuel cell modeling and simulation has aroused much attention recently because it can probe transport and reaction mechanism. In this paper, a computational fuel cell dynamics (CFCD) method was applied to simulate a proton exchange membrane fuel cell (PEMFC) stack for the first time. The air cooling mini fuel cell stack consisted of six cells, in which the active area was 8 cm 2 (2 cm × 4 cm). With reasonable simplification, the computational elements were effectively reduced and allowed a simulation which could be conducted on a personal computer without large-scale parallel computation. The results indicated that the temperature gradient inside the fuel cell stack was determined by the flow rate of the cooling air. If the air flow rate is too low, the stack could not be effectively cooled and the temperature will rise to a range that might cause unstable stack operation.