Downloads & Free Reading Options - Results

The matrix element method utilizes ab initio calculations of probability densities as powerful discriminants for processes of interest in experimental particle physics. The method has already been used successfully at previous and current collider experiments. However, the computational complexity of this method for final states with many particles and degrees of freedom sets it at a disadvantage compared to supervised classification methods such as decision trees, k nearest-neighbour, or neural networks. This note presents a concrete implementation of the matrix element technique using graphics processing units. Due to the intrinsic parallelizability of multidimensional integration, dramatic speedups can be readily achieved, which makes the matrix element technique viable for general usage at collider experiments.

The matrix element method utilizes ab initio calculations of probability densities as powerful discriminants for processes of interest in experimental particle physics. The method has already been used successfully at previous and current collider experiments. However, the computational complexity of this method for final states with many particles and degrees of freedom sets it at a disadvantage compared to supervised classification methods such as decision trees, k nearest-neighbour, or neural networks. This note presents a concrete implementation of the matrix element technique using graphics processing units. Due to the intrinsic parallelizability of multidimensional integration, dramatic speedups can be readily achieved, which makes the matrix element technique viable for general usage at collider experiments.

The matrix element method utilizes ab initio calculations of probability densities as powerful discriminants for processes of interest in experimental particle physics. The method has already been used successfully at previous and current collider experiments. However, the computational complexity of this method for final states with many particles and degrees of freedom sets it at a disadvantage compared to supervised classification methods such as decision trees, k nearest-neighbour, or neural networks. This note presents a concrete implementation of the matrix element technique using graphics processing units. Due to the intrinsic parallelizability of multidimensional integration, dramatic speedups can be readily achieved, which makes the matrix element technique viable for general usage at collider experiments.

The idea of upgrading performance and utility of computer systems by incorporating parallel processing has been around since at least the 1940s. A significant investment in parallel processing in the 1980s and 1990s has led to an abundance of parallel architectures that due to technical constraints at the time had to rely on multi-chip multi-processing. Unfortunately, their impact on mainstream computing was quite limited. 'Tribal lore' suggests the following reason: while programming for parallelism tends to be easy, turning this parallelism into the 'coarse-grained' type, needed to optimize performance for multi-chip multi-processing (with their high coordination overhead), has been quite hard. The mainstream computing paradigm has always relied on serial code. However, commercially successful processors are entering a second decade of near stagnation in the maximum number of instructions they can issue towards a single computational task in one clock. Alas, the multi-decade reliance on advancing the clock is also coming to an end. The PRAM-On-Chip research project at UMD is reviewed. The project is guided by a fact, an old insight, a recent one and a premise. The fact: Billion transistor chips are here, up from less than 30,000 circa 1980. The older insight: Using a very simple parallel computation model, the parallel random access model (PRAM), the parallel algorithms research community succeeded in developing a general-purpose theory of parallel algorithms that was much richer than any competing approach and is second in magnitude only to serial algorithmics. However, since it did not offer an effective abstraction for multi-chip multi-processors this elegant algorithmic theory remained in the ivory towers of theorists. The PRAM-On-Chip insight: The Billion transistor chip era allows for the first time low-overhead on-chip multi-processors so that the PRAM abstraction becomes effective. The premise: Were the architecture component of PRAM-On-Chip feasible in the 1990s, its parallel programming component would become the mainstream standard. In 1988-90 standard algorithms textbooks chose to include significant PRAM chapters (some still have them). Arguably nothing could stand in the way of teaching them to every student at every computer science program. Programming for concurrency/parallelism is quickly becoming an integral part of mainstream computing. Yet, industry and academia leaders in system software and general-purpose application software have maintained a passive posture: their attention tends to focus too much on getting the best performance out of architectures that originated from a hardware centric approach, or very specific applications, and too little on trying to impact the evolving generation of multi-core and/or multi-threaded general-purpose architectures. We argue, perhaps provocatively, that: (i) limiting programming for parallelism to fit hardware-centric architectures imports the epidemic of programmability problems that has plagued parallel computing into mainstream computing, (ii) it is only a matter of time until the industry will seek convergence based on parallel programmability: the difference in the bottom line between a successful and a less successful strategy on parallel programmability will be too big to ignore, (iii) a more assertive position by such leaders is necessary, and (iv) the PRAM-On-Chip approach offers a more balanced way that avoids these problems. URL: http://www.umiacs.umd.edu/~vishkin/XMT ©2005 Microsoft Corporation. All rights reserved.

The use of computers in statistical physics is common because the sheer number of equations that describe the behavior of an entire system particle by particle often makes it impossible to solve them exactly. Monte Carlo methods form a particularly important class of numerical methods for solving problems in statistical physics. Although these methods are simple in principle, their proper use requires a good command of statistical mechanics, as well as considerable computational resources. The aim of this paper is to demonstrate how the usage of widely accessible graphics cards on personal computers can elevate the computing power in Monte Carlo simulations by orders of magnitude, thus allowing live classroom demonstration of phenomena that would otherwise be out of reach. As an example, we use the public goods game on a square lattice where two strategies compete for common resources in a social dilemma situation. We show that the second-order phase transition to an absorbing phase in the system belongs to the directed percolation universality class, and we compare the time needed to arrive at this result by means of the main processor and by means of a suitable graphics card. Parallel computing on graphics processing units has been developed actively during the last decade, to the point where today the learning curve for entry is anything but steep for those familiar with programming. The subject is thus ripe for inclusion in graduate and advanced undergraduate curricula, and we hope that this paper will facilitate this process in the realm of physics education. To that end, we provide a documented source code for an easy reproduction of presented results and for further development of Monte Carlo simulations of similar systems.

The use of computers in statistical physics is common because the sheer number of equations that describe the behavior of an entire system particle by particle often makes it impossible to solve them exactly. Monte Carlo methods form a particularly important class of numerical methods for solving problems in statistical physics. Although these methods are simple in principle, their proper use requires a good command of statistical mechanics, as well as considerable computational resources. The aim of this paper is to demonstrate how the usage of widely accessible graphics cards on personal computers can elevate the computing power in Monte Carlo simulations by orders of magnitude, thus allowing live classroom demonstration of phenomena that would otherwise be out of reach. As an example, we use the public goods game on a square lattice where two strategies compete for common resources in a social dilemma situation. We show that the second-order phase transition to an absorbing phase in the system belongs to the directed percolation universality class, and we compare the time needed to arrive at this result by means of the main processor and by means of a suitable graphics card. Parallel computing on graphics processing units has been developed actively during the last decade, to the point where today the learning curve for entry is anything but steep for those familiar with programming. The subject is thus ripe for inclusion in graduate and advanced undergraduate curricula, and we hope that this paper will facilitate this process in the realm of physics education. To that end, we provide a documented source code for an easy reproduction of presented results and for further development of Monte Carlo simulations of similar systems.

The development of implicit schemes for obtaining steady state solutions to the Euler and Navier-Stokes equations on unstructured grids is outlined. Applications are presented that compare the convergence characteristics of various implicit methods. Next, the development of explicit and implicit schemes to compute unsteady flows on unstructured grids is discussed. Next, the issues involved in parallelizing finite volume schemes on unstructured meshes in an MIMD (multiple instruction/multiple data stream) fashion are outlined. Techniques for partitioning unstructured grids among processors and for extracting parallelism in explicit and implicit solvers are discussed. Finally, some dynamic load balancing ideas, which are useful in adaptive transient computations, are presented.

The aim of the research summarized in this final report was to investigate a class of orthogonal shared-memory architectures and interconnection networks, and to obtain generalized methods for implementing algorithm-based fault tolerance (ABFT) on multiprocessor architectures. We proposed a theory based on orthogonal graphs to represent many well-known interconnection networks such as the binary m-cube, spanning-bus meshes, multistage interconnection networks, etc. A previously proposed multiprocessor architecture called the Orthogonal Multiprocessor (OMP) is also a special case of this method. The simplicity of the graph construction rules permits us to characterize and understand the differences and similarities among networks like the SW-banyan, the baseline network, among others. This opens the way for discovering new structures by studying different possible combinations of the parameters which define orthogonal graphs.

This report describes research performed during the period May 1988- May 1989 under DARPA support. The report contains discussion of four main topics: 1) On-going research on visual navigation, focusing on a system named RAMBO, for the study of robots acting on moving bodies; 2) Development and implementation of parallel algorithms for image processing and computer vision on the Connection Machine and the Butterfly; 3)Development of parallel heuristic search algorithms on the Butterfly that have linear speedup properties over a wide range of problem sizes and machine sizes; and 4) Development of Connection Machine algorithms for matrix operations that are key computational steps in many image processing and computer vision algorithms. This research has resulted in twelve technical reports, and several publications in conferences and workshops. Keywords: Autonomous navigation; Computer vision; Parallel processing; Search.

This report describes research performed during the period May 1988- May 1989 under DARPA support. The report contains discussion of four main topics: 1) On-going research on visual navigation, focusing on a system named RAMBO, for the study of robots acting on moving bodies; 2) Development and implementation of parallel algorithms for image processing and computer vision on the Connection Machine and the Butterfly; 3)Development of parallel heuristic search algorithms on the Butterfly that have linear speedup properties over a wide range of problem sizes and machine sizes; and 4) Development of Connection Machine algorithms for matrix operations that are key computational steps in many image processing and computer vision algorithms. This research has resulted in twelve technical reports, and several publications in conferences and workshops. Keywords: Autonomous navigation; Computer vision; Parallel processing; Search.

Computation of optimal cycle mean in a directed weighted graph has many applications in program analysis, performance verification in particular. In this paper we propose a data-parallel algorithmic solution to the problem and show how the computation of optimal cycle mean can be efficiently accelerated by means of CUDA technology. We show how the problem of computation of optimal cycle mean is decomposed into a sequence of data-parallel graph computation primitives and show how these primitives can be implemented and optimized for CUDA computation. Finally, we report a fivefold experimental speed up on graphs representing models of distributed systems when compared to best sequential algorithms.

Parallel Computing 1989: Volume 12 , Issue Index. Digitized from IA1652917-03 . Previous issue: sim_parallel-computing_1989-08-28_11_3 . Next issue: sim_parallel-computing_1989-10_12_1 .

Parallel Computing 1989: Volume 12 , Issue Index. Digitized from IA1652917-03 . Previous issue: sim_parallel-computing_1989-08-28_11_3 . Next issue: sim_parallel-computing_1989-10_12_1 .

h2Parallel computing in Python: Current state and recent advances/h2 Modern hardware is multi-core. It is crucial for Python to provide high-performance parallelism. This talk will expose to both data-scientists and library developers the current state of affairs and the recent advances for parallel computing with Python. The goal is to help practitioners and developers to make better decisions on this matter. I will first cover how Python can interface with parallelism, from leveraging external parallelism of C-extensions –especially the BLAS family– to Python's multiprocessing and multithreading API. I will touch upon use cases, e.g single vs multi machine, as well as and pros and cons of the various solutions for each use case. Most of these considerations will be backed by benchmarks from the scikit-learn machine learning library. From these low-level interfaces emerged higher-level parallel processing libraries, such as concurrent.futures, joblib and loky (used by dask and scikit-learn) These libraries make it easy for Python programmers to use safe and reliable parallelism in their code. They can even work in more exotic situations, such as interactive sessions, in which Python’s native multiprocessing support tends to fail. I will describe their purpose as well as the canonical use-cases they address. The last part of this talk will focus on the most recent advances in the Python standard library, addressing one of the principal performance bottlenecks of multi-core/multi-machine processing, which is data communication. We will present a new API for shared-memory management between different Python processes, and performance improvements for the serialization of large Python objects ( PEP 574, pickle extensions). These performance improvements will be leveraged by distributed data science frameworks such as dask, ray and pyspark. Please see our speaker release agreement for details: https://ep2019.europython.eu/events/speaker-release-agreement/

h2Parallel computing in Python: Current state and recent advances/h2 Modern hardware is multi-core. It is crucial for Python to provide high-performance parallelism. This talk will expose to both data-scientists and library developers the current state of affairs and the recent advances for parallel computing with Python. The goal is to help practitioners and developers to make better decisions on this matter. I will first cover how Python can interface with parallelism, from leveraging external parallelism of C-extensions –especially the BLAS family– to Python's multiprocessing and multithreading API. I will touch upon use cases, e.g single vs multi machine, as well as and pros and cons of the various solutions for each use case. Most of these considerations will be backed by benchmarks from the scikit-learn machine learning library. From these low-level interfaces emerged higher-level parallel processing libraries, such as concurrent.futures, joblib and loky (used by dask and scikit-learn) These libraries make it easy for Python programmers to use safe and reliable parallelism in their code. They can even work in more exotic situations, such as interactive sessions, in which Python’s native multiprocessing support tends to fail. I will describe their purpose as well as the canonical use-cases they address. The last part of this talk will focus on the most recent advances in the Python standard library, addressing one of the principal performance bottlenecks of multi-core/multi-machine processing, which is data communication. We will present a new API for shared-memory management between different Python processes, and performance improvements for the serialization of large Python objects ( PEP 574, pickle extensions). These performance improvements will be leveraged by distributed data science frameworks such as dask, ray and pyspark. Please see our speaker release agreement for details: https://ep2019.europython.eu/events/speaker-release-agreement/

h2Parallel computing in Python: Current state and recent advances/h2 Modern hardware is multi-core. It is crucial for Python to provide high-performance parallelism. This talk will expose to both data-scientists and library developers the current state of affairs and the recent advances for parallel computing with Python. The goal is to help practitioners and developers to make better decisions on this matter. I will first cover how Python can interface with parallelism, from leveraging external parallelism of C-extensions –especially the BLAS family– to Python's multiprocessing and multithreading API. I will touch upon use cases, e.g single vs multi machine, as well as and pros and cons of the various solutions for each use case. Most of these considerations will be backed by benchmarks from the scikit-learn machine learning library. From these low-level interfaces emerged higher-level parallel processing libraries, such as concurrent.futures, joblib and loky (used by dask and scikit-learn) These libraries make it easy for Python programmers to use safe and reliable parallelism in their code. They can even work in more exotic situations, such as interactive sessions, in which Python’s native multiprocessing support tends to fail. I will describe their purpose as well as the canonical use-cases they address. The last part of this talk will focus on the most recent advances in the Python standard library, addressing one of the principal performance bottlenecks of multi-core/multi-machine processing, which is data communication. We will present a new API for shared-memory management between different Python processes, and performance improvements for the serialization of large Python objects ( PEP 574, pickle extensions). These performance improvements will be leveraged by distributed data science frameworks such as dask, ray and pyspark. Please see our speaker release agreement for details: https://ep2019.europython.eu/events/speaker-release-agreement/

h2Parallel computing in Python: Current state and recent advances/h2 Modern hardware is multi-core. It is crucial for Python to provide high-performance parallelism. This talk will expose to both data-scientists and library developers the current state of affairs and the recent advances for parallel computing with Python. The goal is to help practitioners and developers to make better decisions on this matter. I will first cover how Python can interface with parallelism, from leveraging external parallelism of C-extensions –especially the BLAS family– to Python's multiprocessing and multithreading API. I will touch upon use cases, e.g single vs multi machine, as well as and pros and cons of the various solutions for each use case. Most of these considerations will be backed by benchmarks from the scikit-learn machine learning library. From these low-level interfaces emerged higher-level parallel processing libraries, such as concurrent.futures, joblib and loky (used by dask and scikit-learn) These libraries make it easy for Python programmers to use safe and reliable parallelism in their code. They can even work in more exotic situations, such as interactive sessions, in which Python’s native multiprocessing support tends to fail. I will describe their purpose as well as the canonical use-cases they address. The last part of this talk will focus on the most recent advances in the Python standard library, addressing one of the principal performance bottlenecks of multi-core/multi-machine processing, which is data communication. We will present a new API for shared-memory management between different Python processes, and performance improvements for the serialization of large Python objects ( PEP 574, pickle extensions). These performance improvements will be leveraged by distributed data science frameworks such as dask, ray and pyspark. Please see our speaker release agreement for details: https://ep2019.europython.eu/events/speaker-release-agreement/

Establishing a test-bed for clustered parallel computing is reported along with performance evaluation of various clusters for a number of applications/parallel algorithms.

We introduce process-oriented programming as a natural extension of object-oriented programming for parallel computing. It is based on the observation that every class of an object-oriented language can be instantiated as a process, accessible via a remote pointer. The introduction of process pointers requires no syntax extension, identifies processes with programming objects, and enables processes to exchange information simply by executing remote methods. Process-oriented programming is a high-level language alternative to multithreading, MPI and many other languages, environments and tools currently used for parallel computations. It implements natural object-based parallelism using only minimal syntax extension of existing languages, such as C++ and Python, and has therefore the potential to lead to widespread adoption of parallel programming. We implemented a prototype system for running processes using C++ with MPI and used it to compute a large three-dimensional Fourier transform on a computer cluster built of commodity hardware components. Three-dimensional Fourier transform is a prototype of a data-intensive application with a complex data-access pattern. The process-oriented code is only a few hundred lines long, and attains very high data throughput by achieving massive parallelism and maximizing hardware utilization.

This report summarizes the research performed under Contract DACA76- 88-C-0008 during the period May 1989 - May 1990, the first year of the contract period. The focus of the research program is visual navigation, with an emphasis on the use of massively parallel algorithms to support basic navigational tasks in vision and planning. The first section describes research performed on a project called RAMBO. (RAMBO is an acronym for Robot Acting on Moving BOdies). The project attempts to develop and integrate Connection Machine algorithms for low-level vision, intermediate level vision and visual planning to allow a mobile robot to pursue (in simulation) a moving three dimensional target through space in order to maintain visual contact with points on surface of the target. The next selection describes our past year's work on cross-country navigation. We first describe massively parallel algorithms for route planning in digital terrain maps. Then we describe our research on the problem of filling in range shadows. We discuss why classical interpolation methods are not appropriate for this problem, and present methods. The last section presents brief descriptions of other research projects whose results were reported under this contract during the past year.

This report summarizes the research performed under Contract DACA76- 88-C-0008 during the period May 1989 - May 1990, the first year of the contract period. The focus of the research program is visual navigation, with an emphasis on the use of massively parallel algorithms to support basic navigational tasks in vision and planning. The first section describes research performed on a project called RAMBO. (RAMBO is an acronym for Robot Acting on Moving BOdies). The project attempts to develop and integrate Connection Machine algorithms for low-level vision, intermediate level vision and visual planning to allow a mobile robot to pursue (in simulation) a moving three dimensional target through space in order to maintain visual contact with points on surface of the target. The next selection describes our past year's work on cross-country navigation. We first describe massively parallel algorithms for route planning in digital terrain maps. Then we describe our research on the problem of filling in range shadows. We discuss why classical interpolation methods are not appropriate for this problem, and present methods. The last section presents brief descriptions of other research projects whose results were reported under this contract during the past year.

Volume 2-Computational Fluid dynamics and Wave Propagation; Parallel Computing; Concurrent and Supercomputing; Computational Physics/Computational Chemistry and Evolutionary Systems.

First Class

First Class

polski test turinga gnu marek nowicki

Markov chain Monte Carlo (MCMC) is the predominant tool used in Bayesian parameter estimation for hierarchical models. When the model expands due to an increasing number of hierarchical levels, number of groups at a particular level, or number of observations in each group, a fully Bayesian analysis via MCMC can easily become computationally demanding, even intractable. We illustrate how the steps in an MCMC for hierarchical models are predominantly one of two types: conditionally independent draws or low-dimensional draws based on summary statistics of parameters at higher levels of the hierarchy. Parallel computing can increase efficiency by performing embarrassingly parallel computations for conditionally independent draws and calculating the summary statistics using parallel reductions. During the MCMC algorithm, we record running means and means of squared parameter values to allow convergence diagnosis and posterior inference while avoiding the costly memory transfer bottleneck. We demonstrate the effectiveness of the algorithm on a model motivated by next generation sequencing data, and we release our implementation in R packages fbseq and fbseqCUDA.

Whether or not the problem of finding maximal independent sets (MIS) in hypergraphs is in (R)NC is one of the fundamental problems in the theory of parallel computing. Unlike the well-understood case of MIS in graphs, for the hypergraph problem, our knowledge is quite limited despite considerable work. It is known that the problem is in \emph{RNC} when the edges of the hypergraph have constant size. For general hypergraphs with $n$ vertices and $m$ edges, the fastest previously known algorithm works in time $O(\sqrt{n})$ with $\text{poly}(m,n)$ processors. In this paper we give an EREW PRAM algorithm that works in time $n^{o(1)}$ with $\text{poly}(m,n)$ processors on general hypergraphs satisfying $m \leq n^{\frac{\log^{(2)}n}{8(\log^{(3)}n)^2}}$, where $\log^{(2)}n = \log\log n$ and $\log^{(3)}n = \log\log\log n$. Our algorithm is based on a sampling idea that reduces the dimension of the hypergraph and employs the algorithm for constant dimension hypergraphs as a subroutine.

Whether or not the problem of finding maximal independent sets (MIS) in hypergraphs is in (R)NC is one of the fundamental problems in the theory of parallel computing. Unlike the well-understood case of MIS in graphs, for the hypergraph problem, our knowledge is quite limited despite considerable work. It is known that the problem is in \emph{RNC} when the edges of the hypergraph have constant size. For general hypergraphs with $n$ vertices and $m$ edges, the fastest previously known algorithm works in time $O(\sqrt{n})$ with $\text{poly}(m,n)$ processors. In this paper we give an EREW PRAM algorithm that works in time $n^{o(1)}$ with $\text{poly}(m,n)$ processors on general hypergraphs satisfying $m \leq n^{\frac{\log^{(2)}n}{8(\log^{(3)}n)^2}}$, where $\log^{(2)}n = \log\log n$ and $\log^{(3)}n = \log\log\log n$. Our algorithm is based on a sampling idea that reduces the dimension of the hypergraph and employs the algorithm for constant dimension hypergraphs as a subroutine.

In this paper, we implemented both sequential and parallel version of fractal image compression algorithms using CUDA (Compute Unified Device Architecture) programming model for parallelizing the program in Graphics Processing Unit for medical images, as they are highly similar within the image itself. There are several improvement in the implementation of the algorithm as well. Fractal image compression is based on the self similarity of an image, meaning an image having similarity in majority of the regions. We take this opportunity to implement the compression algorithm and monitor the effect of it using both parallel and sequential implementation. Fractal compression has the property of high compression rate and the dimensionless scheme. Compression scheme for fractal image is of two kind, one is encoding and another is decoding. Encoding is very much computational expensive. On the other hand decoding is less computational. The application of fractal compression to medical images would allow obtaining much higher compression ratios. While the fractal magnification an inseparable feature of the fractal compression would be very useful in presenting the reconstructed image in a highly readable form. However, like all irreversible methods, the fractal compression is connected with the problem of information loss, which is especially troublesome in the medical imaging. A very time consuming encoding pro- cess, which can last even several hours, is another bothersome drawback of the fractal compression.

Conflict of interest is the permanent companion of any population of agents (computational or biological). For that reason, the ability to compromise is of paramount importance, making voting a key element of societal mechanisms. One of the voting procedures most often discussed in the literature and, due to its intuitiveness, also conceptually quite appealing is Charles Dodgson's scoring rule, basically using the respective closeness to being a Condorcet winner for evaluating competing alternatives. In this paper, we offer insights on the practical limits of algorithms computing the exact Dodgson scores from a number of votes. While the problem itself is theoretically intractable, this work proposes and analyses five different solutions which try distinct approaches to practically solve the issue in an effective manner. Additionally, three of the discussed procedures can be run in parallel which has the potential of drastically reducing the problem size.

As the Large Hadron Collider (LHC) continues its upward progression in energy and luminosity towards the planned High-Luminosity LHC (HL-LHC) in 2025, the challenges of the experiments in processing increasingly complex events will also continue to increase. Improvements in computing technologies and algorithms will be a key part of the advances necessary to meet this challenge. Parallel computing techniques, especially those using massively parallel computing (MPC), promise to be a significant part of this effort. In these proceedings, we discuss these algorithms in the specific context of a particularly important problem: the reconstruction of charged particle tracks in the trigger algorithms in an experiment, in which high computing performance is critical for executing the track reconstruction in the available time. We discuss some areas where parallel computing has already shown benefits to the LHC experiments, and also demonstrate how a MPC-based trigger at the CMS experiment could not only improve performance, but also extend the reach of the CMS trigger system to capture events which are currently not practical to reconstruct at the trigger level.

As the Large Hadron Collider (LHC) continues its upward progression in energy and luminosity towards the planned High-Luminosity LHC (HL-LHC) in 2025, the challenges of the experiments in processing increasingly complex events will also continue to increase. Improvements in computing technologies and algorithms will be a key part of the advances necessary to meet this challenge. Parallel computing techniques, especially those using massively parallel computing (MPC), promise to be a significant part of this effort. In these proceedings, we discuss these algorithms in the specific context of a particularly important problem: the reconstruction of charged particle tracks in the trigger algorithms in an experiment, in which high computing performance is critical for executing the track reconstruction in the available time. We discuss some areas where parallel computing has already shown benefits to the LHC experiments, and also demonstrate how a MPC-based trigger at the CMS experiment could not only improve performance, but also extend the reach of the CMS trigger system to capture events which are currently not practical to reconstruct at the trigger level.

We present an efficient distributed memory parallel algorithm for computing connected components in undirected graphs based on Shiloach-Vishkin's PRAM approach. We discuss multiple optimization techniques that reduce communication volume as well as load-balance the algorithm. We also note that the efficiency of the parallel graph connectivity algorithm depends on the underlying graph topology. Particularly for short diameter graph components, we observe that parallel Breadth First Search (BFS) method offers better performance. However, running parallel BFS is not efficient for computing large diameter components or large number of small components. To address this challenge, we employ a heuristic that allows the algorithm to quickly predict the type of the network by computing the degree distribution and follow the optimal hybrid route. Using large graphs with diverse topologies from domains including metagenomics, web crawl, social graph and road networks, we show that our hybrid implementation is efficient and scalable for each of the graph types. Our approach achieves a runtime of 215 seconds using 32K cores of Cray XC30 for a metagenomic graph with over 50 billion edges. When compared against the previous state-of-the-art method, we see performance improvements up to 24x.

Recently, networked and cluster computation have become very popular. This paper is an introduction to a new C based parallel language for architecture-adaptive programming, aCe C. The primary purpose of aCe (Architecture-adaptive Computing Environment) is to encourage programmers to implement applications on parallel architectures by providing them the assurance that future architectures will be able to run their applications with a minimum of modification. A secondary purpose is to encourage computer architects to develop new types of architectures by providing an easily implemented software development environment and a library of test applications. This new language should be an ideal tool to teach parallel programming. In this paper, we will focus on some fundamental features of aCe C.

Cluster-based data-parallel frameworks such as MapReduce, Hadoop, and Dryad are increasingly popular for a large class of compute-intensive tasks. Although such systems are designed for large-scale clusters, they also offer a convenient and accessible route to data-parallel programming for small-scale clusters. This potentially allows applications traditionally targeted at supercomputers or remote server farms, such as sophisticated video processing, to be deployed in a small-scale ad-hoc fashion by aggregating the servers and workstations in the home or office network. The default scheduling algorithms of these frameworks perform well at scale, but are significantly less optimal in a small (3-10 machine) cluster environment where nodes have widely differing performance characteristics. To make effective use of an ad-hoc cluster, we require a 'planner' rather than a scheduler that takes account of the predicted resource consumption by each vertex in the dataflow graph and the heterogeneity of the available hardware. In this talk I will describe our enhancements to DryadLINQ and Dryad for ad-hoc clusters. We have integrated a constraint-based planner that maps the dataflow graph generated by the DryadLINQ compiler onto the cluster. The planner makes use of DryadLINQ operator performance models that are constructed from low-level traces of vertex executions. The performance models abstract the behaviour of each vertex in sufficient detail to predict the bottleneck resource, which can change during vertex execution, on different hardware and with different sizes of input. Experimental evaluation shows reasonable predictive accuracy and good performance gains for parallel jobs on ad-hoc clusters. ©2009 Microsoft Corporation. All rights reserved.

Parallel Computing 1990: Volume 14 , Issue Index. Digitized from IA1652917-03 . Previous issue: sim_parallel-computing_1990-03_13_3 . Next issue: sim_parallel-computing_1990-05_14_1 .

Scalable Parallel Computing on Many/Multicore Systems This set of lectures will review the application and programming model issues that will one must address when one gets chips with 32-1024 cores and “scalable” approaches will be needed to make good use of such systems. We will not discuss bit-level and instruction-level parallelism i.e. what happens on a possibly special purpose core, even though this is clearly important. We will use science and engineering applications to drive the discussion as we have substantial experience in these cases but we are interested in the lessons for commodity client and server applications that will be broadly used 5-10 years from now. We start with simple applications and algorithms from a variety of fields and identify features that makes it “obvious” that “all” science and engineering run well and scalably in parallel. We explain why unfortunately it is equally obvious that there is no straightforward way of expressing this parallelism. Parallel hardware architectures are described in enough detail to understand performance and algorithm issues and need for cross-architecture compatibility; however in these lectures we will just be users of hardware. We must understand what features of multicore chips we can and should exploit. We can explicitly use the shared memory between cores or just enjoy its implications for very fast inter-core control and communication linkage. We note that parallel algorithm research is hugely successful although this success has reduced activity in an area that deserves new attention for the next generation of architectures. The parallel software environment is discussed at several levels including programming paradigm, runtime and operating system. The importance of libraries, templates, kernels (dwarfs) and benchmarks is stressed. The programming environment has various tools including compilers with possible parallelism hints like OpenMP; tuners like Atlas; messaging models; parallelism and distribution support as in HPF, HPCS Languages, co-array Fortran, Global arrays and UPC. We also discuss the relevance of important general ideas like object-oriented paradigms (as in for example Charm++), functional languages and Software Transactional Memories. Streaming, pipelining, co-ordination, services and workflow are placed in context. Examples discussed in the last category include CCR/DSS from Microsoft and the Common Component Architecture CCA from DoE. Domain Specific environments like Matlab and Mathematica are important as there is no universal silver programming bullet; one will need interoperable focused environments. We discuss performance analysis including speedup, efficiency, scaled speedup and Amdahl's law. We show how to relate performance to algorithm/application structure and the hardware characteristics. Applications will not get scalable parallelism accidentally but only if there is an understandable performance model. We review some of the many pitfalls for both performance and correctness; these include deadlocks, race conditions, nodes that are busy doing something else and the difficulty of second-guessing automatic parallelization methods. We describe the formal sources of overhead; load imbalance and the communication/control overhead and the ways to reduce them. We relate the blocking used in caching to that used in parallel computation. We note that load balancing was the one part of parallel computing that was easier than expected. We will mix both simple idealized applications with “real problems” noting that usually it is simple problems that are the hardest as they have poor computation to control/communication ratios. We will explain the parallelism in several application classes including for science and engineering: Finite Difference, Finite Elements, FFT/Spectral, Meshes of all sorts, Particle Dynamics, Particle-Mesh, and Monte Carlo methods. Some applications like Image Processing, Graphics, Cryptography and Media coding/decoding have features similar to well understood science and engineering problems. We emphasize that nearly all applications are built hierarchically from more “basic applications” with a variety of different structures and natural programming models. Such application composition (co-ordination, workflow) must be supported with a common run-time. We contrast the difference between decomposition needed in most “basic parallel applications” to the composition supported in workflow and Web 2.0 mashups. Looking at broader application classes we should cover; Internet applications and services; artificial intelligence, optimization, machine learning; divide and conquer algorithms; tree structured searches like computer chess; applications generated by the data deluge including access, search, and the assimilation of data into simulations. There has been a growing interest in Complex Systems whether it be for critical infrastructure like energy and transportation, the Internet itself, commodity gaming, computational epidemiology or the original war games. We expect Discrete Event Simulations (such as DoD’s High Level Architecture HLA) to grow in importance as they naturally describe complex systems and because they can clearly benefit from multicore architectures. We will discuss the innovative Sequential Dynamical Systems approach used in the EpiSims, TransSims and other critical infrastructure simulation environments. In all applications we need to identify the intrinsic parallelism and the degrees of freedom that can be parallelized and distinguish small parallelism (local to core) compared to large parallelism (scalable across many cores) We will not forget many critical non-technical Issues including “Who programs – everybody or just the marine corps?”; “The market for science and engineering was small but it will be large for general multicore”; “The program exists and can’t be changed”; “What features will next hardware/software release support and how should I protect myself from change?” We will summarize lessons and relate them to application and programming model categories. In last lecture or at end of all lectures we encourage the audience to bring their own multicore application or programming model so we can discuss examples that interest you. ©2007 Microsoft Corporation. All rights reserved.

Scalable Parallel Computing on Many/Multicore Systems This set of lectures will review the application and programming model issues that will one must address when one gets chips with 32-1024 cores and “scalable” approaches will be needed to make good use of such systems. We will not discuss bit-level and instruction-level parallelism i.e. what happens on a possibly special purpose core, even though this is clearly important. We will use science and engineering applications to drive the discussion as we have substantial experience in these cases but we are interested in the lessons for commodity client and server applications that will be broadly used 5-10 years from now. We start with simple applications and algorithms from a variety of fields and identify features that makes it “obvious” that “all” science and engineering run well and scalably in parallel. We explain why unfortunately it is equally obvious that there is no straightforward way of expressing this parallelism. Parallel hardware architectures are described in enough detail to understand performance and algorithm issues and need for cross-architecture compatibility; however in these lectures we will just be users of hardware. We must understand what features of multicore chips we can and should exploit. We can explicitly use the shared memory between cores or just enjoy its implications for very fast inter-core control and communication linkage. We note that parallel algorithm research is hugely successful although this success has reduced activity in an area that deserves new attention for the next generation of architectures. The parallel software environment is discussed at several levels including programming paradigm, runtime and operating system. The importance of libraries, templates, kernels (dwarfs) and benchmarks is stressed. The programming environment has various tools including compilers with possible parallelism hints like OpenMP; tuners like Atlas; messaging models; parallelism and distribution support as in HPF, HPCS Languages, co-array Fortran, Global arrays and UPC. We also discuss the relevance of important general ideas like object-oriented paradigms (as in for example Charm++), functional languages and Software Transactional Memories. Streaming, pipelining, co-ordination, services and workflow are placed in context. Examples discussed in the last category include CCR/DSS from Microsoft and the Common Component Architecture CCA from DoE. Domain Specific environments like Matlab and Mathematica are important as there is no universal silver programming bullet; one will need interoperable focused environments. We discuss performance analysis including speedup, efficiency, scaled speedup and Amdahl's law. We show how to relate performance to algorithm/application structure and the hardware characteristics. Applications will not get scalable parallelism accidentally but only if there is an understandable performance model. We review some of the many pitfalls for both performance and correctness; these include deadlocks, race conditions, nodes that are busy doing something else and the difficulty of second-guessing automatic parallelization methods. We describe the formal sources of overhead; load imbalance and the communication/control overhead and the ways to reduce them. We relate the blocking used in caching to that used in parallel computation. We note that load balancing was the one part of parallel computing that was easier than expected. We will mix both simple idealized applications with “real problems” noting that usually it is simple problems that are the hardest as they have poor computation to control/communication ratios. We will explain the parallelism in several application classes including for science and engineering: Finite Difference, Finite Elements, FFT/Spectral, Meshes of all sorts, Particle Dynamics, Particle-Mesh, and Monte Carlo methods. Some applications like Image Processing, Graphics, Cryptography and Media coding/decoding have features similar to well understood science and engineering problems. We emphasize that nearly all applications are built hierarchically from more “basic applications” with a variety of different structures and natural programming models. Such application composition (co-ordination, workflow) must be supported with a common run-time. We contrast the difference between decomposition needed in most “basic parallel applications” to the composition supported in workflow and Web 2.0 mashups. Looking at broader application classes we should cover; Internet applications and services; artificial intelligence, optimization, machine learning; divide and conquer algorithms; tree structured searches like computer chess; applications generated by the data deluge including access, search, and the assimilation of data into simulations. There has been a growing interest in Complex Systems whether it be for critical infrastructure like energy and transportation, the Internet itself, commodity gaming, computational epidemiology or the original war games. We expect Discrete Event Simulations (such as DoD’s High Level Architecture HLA) to grow in importance as they naturally describe complex systems and because they can clearly benefit from multicore architectures. We will discuss the innovative Sequential Dynamical Systems approach used in the EpiSims, TransSims and other critical infrastructure simulation environments. In all applications we need to identify the intrinsic parallelism and the degrees of freedom that can be parallelized and distinguish small parallelism (local to core) compared to large parallelism (scalable across many cores) We will not forget many critical non-technical Issues including “Who programs – everybody or just the marine corps?”; “The market for science and engineering was small but it will be large for general multicore”; “The program exists and can’t be changed”; “What features will next hardware/software release support and how should I protect myself from change?” We will summarize lessons and relate them to application and programming model categories. In last lecture or at end of all lectures we encourage the audience to bring their own multicore application or programming model so we can discuss examples that interest you. ©2007 Microsoft Corporation. All rights reserved.

An O(log sup 2 N) parallel algorithm is presented for computing the eigenvalues of a symmetric tridiagonal matrix using a parallel algorithm for computing the zeros of the characteristic polynomial. The method is based on a quadratic recurrence in which the characteristic polynomial is constructed on a binary tree from polynomials whose degree doubles at each level. Intervals that contain exactly one zero are determined by the zeros of polynomials at the previous level which ensures that different processors compute different zeros. The exact behavior of the polynomials at the interval endpoints is used to eliminate the usual problems induced by finite precision arithmetic.

We present an evaluation of a simulated cosmic ray shower, based on {\sc geant4} and {\sc top-c}, which tracks all the particles in the shower. {\sc top-c} (Task Oriented Parallel C) provides a framework for parallel algorithm development which makes tractable the problem of following each particle. This method is compared with a simulation program which employs the Hillas thinning algorithm.

Offline software using TCP/IP sockets to distribute particle physics events to multiple UNIX/RISC workstations is described. A modular, building block approach was taken, which allowed tailoring to solve specific tasks efficiently and simply as they arose. The modest, initial cost was having to learn about sockets for interprocess communication. This multiprocessor management software has been used to control the reconstruction of eight billion raw data events from Fermilab Experiment E791.

The Voronoi diagram is a widely used data structure. The theory of algorithms for computing Euclidean Voronoi diagrams of point sites is rich and useful, with several different and important algorithms. However, this theory has been quite steady during the last few decades in the sense that new algorithms have not entered the game. In addition, most of the known algorithms are sequential in nature and hence cast inherent difficulties on the possibility to compute the diagram in parallel. This paper presents a new and simple algorithm which enables the (combinatorial) computation of the diagram. The algorithm is significantly different from previous ones and some of the involved concepts in it are in the spirit of linear programming and optics. Parallel implementation is naturally supported since each Voronoi cell can be computed independently of the other cells. A new combinatorial structure for representing the cells (and any convex polytope) is described along the way and the computation of the induced Delaunay graph is obtained almost automatically.

We present techniques to parallelize membership tests for Deterministic Finite Automata (DFAs). Our method searches arbitrary regular expressions by matching multiple bytes in parallel using speculation. We partition the input string into chunks, match chunks in parallel, and combine the matching results. Our parallel matching algorithm exploits structural DFA properties to minimize the speculative overhead. Unlike previous approaches, our speculation is failure-free, i.e., (1) sequential semantics are maintained, and (2) speed-downs are avoided altogether. On architectures with a SIMD gather-operation for indexed memory loads, our matching operation is fully vectorized. The proposed load-balancing scheme uses an off-line profiling step to determine the matching capacity of each par- ticipating processor. Based on matching capacities, DFA matches are load-balanced on inhomogeneous parallel architectures such as cloud computing environments. We evaluated our speculative DFA membership test for a representative set of benchmarks from the Perl-compatible Regular Expression (PCRE) library and the PROSITE protein database. Evaluation was conducted on a 4 CPU (40 cores) shared-memory node of the Intel Manycore Testing Lab (Intel MTL), on the Intel AVX2 SDE simulator for 8-way fully vectorized SIMD execution, and on a 20-node (288 cores) cluster on the Amazon EC2 computing cloud.

We present techniques to parallelize membership tests for Deterministic Finite Automata (DFAs). Our method searches arbitrary regular expressions by matching multiple bytes in parallel using speculation. We partition the input string into chunks, match chunks in parallel, and combine the matching results. Our parallel matching algorithm exploits structural DFA properties to minimize the speculative overhead. Unlike previous approaches, our speculation is failure-free, i.e., (1) sequential semantics are maintained, and (2) speed-downs are avoided altogether. On architectures with a SIMD gather-operation for indexed memory loads, our matching operation is fully vectorized. The proposed load-balancing scheme uses an off-line profiling step to determine the matching capacity of each par- ticipating processor. Based on matching capacities, DFA matches are load-balanced on inhomogeneous parallel architectures such as cloud computing environments. We evaluated our speculative DFA membership test for a representative set of benchmarks from the Perl-compatible Regular Expression (PCRE) library and the PROSITE protein database. Evaluation was conducted on a 4 CPU (40 cores) shared-memory node of the Intel Manycore Testing Lab (Intel MTL), on the Intel AVX2 SDE simulator for 8-way fully vectorized SIMD execution, and on a 20-node (288 cores) cluster on the Amazon EC2 computing cloud.

We present techniques to parallelize membership tests for Deterministic Finite Automata (DFAs). Our method searches arbitrary regular expressions by matching multiple bytes in parallel using speculation. We partition the input string into chunks, match chunks in parallel, and combine the matching results. Our parallel matching algorithm exploits structural DFA properties to minimize the speculative overhead. Unlike previous approaches, our speculation is failure-free, i.e., (1) sequential semantics are maintained, and (2) speed-downs are avoided altogether. On architectures with a SIMD gather-operation for indexed memory loads, our matching operation is fully vectorized. The proposed load-balancing scheme uses an off-line profiling step to determine the matching capacity of each par- ticipating processor. Based on matching capacities, DFA matches are load-balanced on inhomogeneous parallel architectures such as cloud computing environments. We evaluated our speculative DFA membership test for a representative set of benchmarks from the Perl-compatible Regular Expression (PCRE) library and the PROSITE protein database. Evaluation was conducted on a 4 CPU (40 cores) shared-memory node of the Intel Manycore Testing Lab (Intel MTL), on the Intel AVX2 SDE simulator for 8-way fully vectorized SIMD execution, and on a 20-node (288 cores) cluster on the Amazon EC2 computing cloud.

We present techniques to parallelize membership tests for Deterministic Finite Automata (DFAs). Our method searches arbitrary regular expressions by matching multiple bytes in parallel using speculation. We partition the input string into chunks, match chunks in parallel, and combine the matching results. Our parallel matching algorithm exploits structural DFA properties to minimize the speculative overhead. Unlike previous approaches, our speculation is failure-free, i.e., (1) sequential semantics are maintained, and (2) speed-downs are avoided altogether. On architectures with a SIMD gather-operation for indexed memory loads, our matching operation is fully vectorized. The proposed load-balancing scheme uses an off-line profiling step to determine the matching capacity of each par- ticipating processor. Based on matching capacities, DFA matches are load-balanced on inhomogeneous parallel architectures such as cloud computing environments. We evaluated our speculative DFA membership test for a representative set of benchmarks from the Perl-compatible Regular Expression (PCRE) library and the PROSITE protein database. Evaluation was conducted on a 4 CPU (40 cores) shared-memory node of the Intel Manycore Testing Lab (Intel MTL), on the Intel AVX2 SDE simulator for 8-way fully vectorized SIMD execution, and on a 20-node (288 cores) cluster on the Amazon EC2 computing cloud.

We present techniques to parallelize membership tests for Deterministic Finite Automata (DFAs). Our method searches arbitrary regular expressions by matching multiple bytes in parallel using speculation. We partition the input string into chunks, match chunks in parallel, and combine the matching results. Our parallel matching algorithm exploits structural DFA properties to minimize the speculative overhead. Unlike previous approaches, our speculation is failure-free, i.e., (1) sequential semantics are maintained, and (2) speed-downs are avoided altogether. On architectures with a SIMD gather-operation for indexed memory loads, our matching operation is fully vectorized. The proposed load-balancing scheme uses an off-line profiling step to determine the matching capacity of each par- ticipating processor. Based on matching capacities, DFA matches are load-balanced on inhomogeneous parallel architectures such as cloud computing environments. We evaluated our speculative DFA membership test for a representative set of benchmarks from the Perl-compatible Regular Expression (PCRE) library and the PROSITE protein database. Evaluation was conducted on a 4 CPU (40 cores) shared-memory node of the Intel Manycore Testing Lab (Intel MTL), on the Intel AVX2 SDE simulator for 8-way fully vectorized SIMD execution, and on a 20-node (288 cores) cluster on the Amazon EC2 computing cloud.

A previous knowledge of the domains of dependence of an Hamilton Jacobi equation can be useful in its study and approximation. Information of this nature are, in general, difficult to obtain directly from the data of the problem. In this paper we introduce formally the concept of independent sub-domains discussing their main properties and we provide a constructive implicit representation formula. Using such results we propose an algorithm for the approximation of these sets that will be shown to be relevant in parallel computing of the solution.