Autonomic scheduling of tasks from data parallel patterns to CPU/GPU core mixes

We propose a methodology for optimizing the execution of data parallel (sub-)tasks on CPU and GPU cores of the same heterogeneous architecture. The methodology is based on two main components: i) an analytical performance model for scheduling tasks among CPU and GPU cores, such that the global execution time of the overall data parallel pattern is optimized; and ii) an autonomic module which uses the analytical performance model to implement the data parallel computations in a completely autonomic way, requiring no programmer intervention to optimize the computation across CPU and GPU cores. The analytical performance model uses a small set of simple parameters to devise a partitioning-between CPU and GPU cores-of the tasks derived from structured data parallel patterns/algorithmic skeletons. The model takes into account both hardware related and application dependent parameters. It computes the percentage of tasks to be executed on CPU and GPU cores such that both kinds of cores are exploited and performance figures are optimized. The autonomic module, implemented in FastFlow, executes a generic map (reduce) data parallel pattern scheduling part of the tasks to the GPU and part to CPU cores so as to achieve optimal execution time. Experimental results on state-of-the-art CPU/GPU architectures are shown that assess both performance model properties and autonomic module effectiveness. © 2013 IEEE.

Taking Baby Steps in Molecular Logic-based Computation

Molecular logic-based computation is a broad umbrella covering molecular sensors at its simplest level and logic gate arrays involving steadily increasing levels of parallel and serial integration. The fluorescent PET(photoinduced electron transfer) switching principle remains a loyal servant of this entire field. Applications arise from the convenient operation of molecular information processors in very small spaces.

Pool Evolution: A Parallel Pattern for Evolutionary and Symbolic Computing

We introduce a new parallel pattern derived from a specific application domain and show how it turns out to have application beyond its domain of origin. The pool evolution pattern models the parallel evolution of a population subject to mutations and evolving in such a way that a given fitness function is optimized. The pattern has been demonstrated to be suitable for capturing and modeling the parallel patterns underpinning various evolutionary algorithms, as well as other parallel patterns typical of symbolic computation. In this paper we introduce the pattern, we discuss its implementation on modern multi/many core architectures and finally present experimental results obtained with FastFlow and Erlang implementations to assess its feasibility and scalability.

A framework for the development of parallel and distributed real-time embedded systems

Embedded real-time applications increasingly present high computation requirements, which need to be completed within specific deadlines, but that present highly variable patterns, depending on the set of data available in a determined instant. The current trend to provide parallel processing in the embedded domain allows providing higher processing power; however, it does not address the variability in the processing pattern. Dimensioning each device for its worst-case scenario implies lower average utilization, and increased available, but unusable, processing in the overall system. A solution for this problem is to extend the parallel execution of the applications, allowing networked nodes to distribute the workload, on peak situations, to neighbour nodes. In this context, this report proposes a framework to develop parallel and distributed real-time embedded applications, transparently using OpenMP and Message Passing Interface (MPI), within a programming model based on OpenMP. The technical report also devises an integrated timing model, which enables the structured reasoning on the timing behaviour of these hybrid architectures.

ACTORS: A Model of Concurrent Computation in Distributed Systems

A foundational model of concurrency is developed in this thesis. We examine issues in the design of parallel systems and show why the actor model is suitable for exploiting large-scale parallelism. Concurrency in actors is constrained only by the availability of hardware resources and by the logical dependence inherent in the computation. Unlike dataflow and functional programming, however, actors are dynamically reconfigurable and can model shared resources with changing local state. Concurrency is spawned in actors using asynchronous message-passing, pipelining, and the dynamic creation of actors. This thesis deals with some central issues in distributed computing. Specifically, problems of divergence and deadlock are addressed. For example, actors permit dynamic deadlock detection and removal. The problem of divergence is contained because independent transactions can execute concurrently and potentially infinite processes are nevertheless available for interaction.

Thread Scheduling Mechanisms for Multiple-Context Parallel Processors

Scheduling tasks to efficiently use the available processor resources is crucial to minimizing the runtime of applications on shared-memory parallel processors. One factor that contributes to poor processor utilization is the idle time caused by long latency operations, such as remote memory references or processor synchronization operations. One way of tolerating this latency is to use a processor with multiple hardware contexts that can rapidly switch to executing another thread of computation whenever a long latency operation occurs, thus increasing processor utilization by overlapping computation with communication. Although multiple contexts are effective for tolerating latency, this effectiveness can be limited by memory and network bandwidth, by cache interference effects among the multiple contexts, and by critical tasks sharing processor resources with less critical tasks. This thesis presents techniques that increase the effectiveness of multiple contexts by intelligently scheduling threads to make more efficient use of processor pipeline, bandwidth, and cache resources. This thesis proposes thread prioritization as a fundamental mechanism for directing the thread schedule on a multiple-context processor. A priority is assigned to each thread either statically or dynamically and is used by the thread scheduler to decide which threads to load in the contexts, and to decide which context to switch to on a context switch. We develop a multiple-context model that integrates both cache and network effects, and shows how thread prioritization can both maintain high processor utilization, and limit increases in critical path runtime caused by multithreading. The model also shows that in order to be effective in bandwidth limited applications, thread prioritization must be extended to prioritize memory requests. We show how simple hardware can prioritize the running of threads in the multiple contexts, and the issuing of requests to both the local memory and the network. Simulation experiments show how thread prioritization is used in a variety of applications. Thread prioritization can improve the performance of synchronization primitives by minimizing the number of processor cycles wasted in spinning and devoting more cycles to critical threads. Thread prioritization can be used in combination with other techniques to improve cache performance and minimize cache interference between different working sets in the cache. For applications that are critical path limited, thread prioritization can improve performance by allowing processor resources to be devoted preferentially to critical threads. These experimental results show that thread prioritization is a mechanism that can be used to implement a wide range of scheduling policies.

Proposal of a parallel architecture for a motion detection algorithm

This paper proposes a parallel architecture for estimation of the motion of an underwater robot. It is well known that image processing requires a huge amount of computation, mainly at low-level processing where the algorithms are dealing with a great number of data. In a motion estimation algorithm, correspondences between two images have to be solved at the low level. In the underwater imaging, normalised correlation can be a solution in the presence of non-uniform illumination. Due to its regular processing scheme, parallel implementation of the correspondence problem can be an adequate approach to reduce the computation time. Taking into consideration the complexity of the normalised correlation criteria, a new approach using parallel organisation of every processor from the architecture is proposed

Scalability of efficient parallel K-Means

Clustering is defined as the grouping of similar items in a set, and is an important process within the field of data mining. As the amount of data for various applications continues to increase, in terms of its size and dimensionality, it is necessary to have efficient clustering methods. A popular clustering algorithm is K-Means, which adopts a greedy approach to produce a set of K-clusters with associated centres of mass, and uses a squared error distortion measure to determine convergence. Methods for improving the efficiency of K-Means have been largely explored in two main directions. The amount of computation can be significantly reduced by adopting a more efficient data structure, notably a multi-dimensional binary search tree (KD-Tree) to store either centroids or data points. A second direction is parallel processing, where data and computation loads are distributed over many processing nodes. However, little work has been done to provide a parallel formulation of the efficient sequential techniques based on KD-Trees. Such approaches are expected to have an irregular distribution of computation load and can suffer from load imbalance. This issue has so far limited the adoption of these efficient K-Means techniques in parallel computational environments. In this work, we provide a parallel formulation for the KD-Tree based K-Means algorithm and address its load balancing issues.

Dynamic load balancing in parallel KD-tree k-means

One among the most influential and popular data mining methods is the k-Means algorithm for cluster analysis. Techniques for improving the efficiency of k-Means have been largely explored in two main directions. The amount of computation can be significantly reduced by adopting geometrical constraints and an efficient data structure, notably a multidimensional binary search tree (KD-Tree). These techniques allow to reduce the number of distance computations the algorithm performs at each iteration. A second direction is parallel processing, where data and computation loads are distributed over many processing nodes. However, little work has been done to provide a parallel formulation of the efficient sequential techniques based on KD-Trees. Such approaches are expected to have an irregular distribution of computation load and can suffer from load imbalance. This issue has so far limited the adoption of these efficient k-Means variants in parallel computing environments. In this work, we provide a parallel formulation of the KD-Tree based k-Means algorithm for distributed memory systems and address its load balancing issue. Three solutions have been developed and tested. Two approaches are based on a static partitioning of the data set and a third solution incorporates a dynamic load balancing policy.

Hierarchical composite regular parallel architecture

The design space of emerging heterogenous multi-core architectures with re-configurability element makes it feasible to design mixed fine-grained and coarse-grained parallel architectures. This paper presents a hierarchical composite array design which extends the curret design space of regular array design by combining a sequence of transformations. This technique is applied to derive a new design of a pipelined parallel regular array with different dataflow between phases of computation.

A comparative study of Java and C performance in two large-scale parallel applications

In the 1990s the Message Passing Interface Forum defined MPI bindings for Fortran, C, and C++. With the success of MPI these relatively conservative languages have continued to dominate in the parallel computing community. There are compelling arguments in favour of more modern languages like Java. These include portability, better runtime error checking, modularity, and multi-threading. But these arguments have not converted many HPC programmers, perhaps due to the scarcity of full-scale scientific Java codes, and the lack of evidence for performance competitive with C or Fortran. This paper tries to redress this situation by porting two scientific applications to Java. Both of these applications are parallelized using our thread-safe Java messaging system—MPJ Express. The first application is the Gadget-2 code, which is a massively parallel structure formation code for cosmological simulations. The second application uses the finite-domain time-difference method for simulations in the area of computational electromagnetics. We evaluate and compare the performance of the Java and C versions of these two scientific applications, and demonstrate that the Java codes can achieve performance comparable with legacy applications written in conventional HPC languages. Copyright © 2009 John Wiley & Sons, Ltd.

A parallel convolutional coder including embedded puncturing with application to consumer devices

As consumers demand more functionality) from their electronic devices and manufacturers supply the demand then electrical power and clock requirements tend to increase, however reassessing system architecture can fortunately lead to suitable counter reductions. To maintain low clock rates and therefore reduce electrical power, this paper presents a parallel convolutional coder for the transmit side in many wireless consumer devices. The coder accepts a parallel data input and directly computes punctured convolutional codes without the need for a separate puncturing operation while the coded bits are available at the output of the coder in a parallel fashion. Also as the computation is in parallel then the coder can be clocked at 7 times slower than the conventional shift-register based convolutional coder (using DVB 7/8 rate). The presented coder is directly relevant to the design of modern low-power consumer devices

Non-uniform data distribution for communication-efficient parallel clustering

Global communicationrequirements andloadimbalanceof someparalleldataminingalgorithms arethe major obstacles to exploitthe computational power of large-scale systems. This work investigates how non-uniform data distributions can be exploited to remove the global communication requirement and to reduce the communication costin parallel data mining algorithms and, in particular, in the k-means algorithm for cluster analysis. In the straightforward parallel formulation of the k-means algorithm, data and computation loads are uniformly distributed over the processing nodes. This approach has excellent load balancing characteristics that may suggest it could scale up to large and extreme-scale parallel computing systems. However, at each iteration step the algorithm requires a global reduction operationwhichhinders thescalabilityoftheapproach.Thisworkstudiesadifferentparallelformulation of the algorithm where the requirement of global communication is removed, while maintaining the same deterministic nature ofthe centralised algorithm. The proposed approach exploits a non-uniform data distribution which can be either found in real-world distributed applications or can be induced by means ofmulti-dimensional binary searchtrees. The approachcanalso be extended to accommodate an approximation error which allows a further reduction ofthe communication costs. The effectiveness of the exact and approximate methods has been tested in a parallel computing system with 64 processors and in simulations with 1024 processing element

Parallel pipelined histogram architecture via c-slow retiming

A parallel pipelined array of cells suitable for realtime computation of histograms is proposed. The cell architecture builds on previous work to now allow operating on a stream of data at 1 pixel per clock cycle. This new cell is more suitable for interfacing to camera sensors or to microprocessors of 8-bit data buses which are common in consumer digital cameras. Arrays using the new proposed cells are obtained via C-slow retiming techniques and can be clocked at a 65% faster frequency than previous arrays. This achieves over 80% of the performance of two-pixel per clock cycle parallel pipelined arrays.

A parallel quantum histogram architecture

A parallel formulation of an algorithm for the histogram computation of n data items using an on-the-fly data decomposition and a novel quantum-like representation (QR) is developed. The QR transformation separates multiple data read operations from multiple bin update operations thereby making it easier to bind data items into their corresponding histogram bins. Under this model the steps required to compute the histogram is n/s + t steps, where s is a speedup factor and t is associated with pipeline latency. Here, we show that an overall speedup factor, s, is available for up to an eightfold acceleration. Our evaluation also shows that each one of these cells requires less area/time complexity compared to similar proposals found in the literature.