Parallel Processing Letters

Speculative parallelism refers to searching in parallel for a solution, such as finding a pattern in a data base, where finding the first solution terminates the whole parallel process. Different performance prediction methods are required as compared to traditional parallelism. In this paper we introduce an analytical approach to predict the execution time distribution of data-dependent parallel programs that feature N-ary and binary speculative parallel compositions. The method is based on the use of statistical moments which allows program execution time distribution to be approximated at O(1) solution complexity. Measurement results for synthetic distributions indicate an accuracy that lies in the percent range while for empirical distributions on internet search engines the prediction accuracy is acceptable, provided sufficient workload unimodality.

Partial redundancy elimination (PRE) is a key technology for modern compilers. However traditional approaches are conservative and fail to exploit many opportunities for optimization. New PRE approaches which greatly increase the number of eliminated redundancies have been developed. However, they either cause the code size to explode or they cannot handle statements with side-effects. In this paper we describe a predicated partial redundancy elimination (PPRE) approach which can potentially remove all partial redundancies. To avoid performance overheads caused by predication, PPRE is applied selectively based on a cost model. The cost analysis presented in the paper utilizes probabilistic data-flow information to decide whether PPRE is profitable for each instance of a partially redundant computation. Refinements of the basic PPRE transformation are described in detail. In contrast to some other approaches our transformation is strictly semantics preserving.

Notification services mediate between information publishers and consumers that wish to subscribe to periodic updates. In many cases, however, there is a mismatch between the dissemination of these updates and the delivery preferences of the consumer, often in terms of frequency of delivery, quality, etc. In this paper, we present an automated negotiation engine that identifies mutually acceptable terms; we study its performance, and discuss its application to a Grid notification service. We also demonstrate how the negotiation engine enables users to control the Quality of Service levels they require.

When designing and implementing highly efficient scientific applications for parallel computers such as clusters of workstations, it is inevitable to consider and to optimize the single-CPU performance of the codes. For this purpose, it is particularly important that the codes respect the hierarchical memory designs that computer architects employ in order to hide the effects of the growing gap between CPU performance and main memory speed. In this article, we present techniques to enhance the single-CPU efficiency of lattice Boltzmann methods which are commonly used in computational fluid dynamics. We show various performance results for both 2D and 3D codes in order to emphasize the effectiveness of our optimization techniques.

BlueGene/L is a 65,536-compute node massively parallel supercomputer, built using system-on-a-chip integration and a cellular architecture. BlueGene/L represents a major challenge for parallel system software, particularly in the areas of scalability, maintainability, and usability. In this paper, we present the organization of the BlueGene/L system software, with emphasis on the features that address those challenges. The system software was developed in parallel with the hardware, relying on an architecturally accurate simulator of the machine. We validated this approach by demonstrating a working system software stack and high performance on real parallel applications just a few weeks after first hardware availability.

Skeletons are reusable, parameterized program components with well-defined semantics and pre-packaged efficient parallel implementation. This paper develops a new, provably cost-optimal implementation of the DS (double-scan) skeleton for programming divide-and-conquer algorithms. Our implementation is based on a novel data structure called plist (pointed list); implementation's performance is estimated using an analytical model. We demonstrate the use of the DS skeleton for parallelizing a tridiagonal system solver and report experimental results for its MPI implementation on a Cray T3E and a Linux cluster: they confirm the performance improvement achieved by the cost-optimal implementation and demonstrate its good predictability by our performance model.

Distributed Video-on-Demand (DVoD) systems are proposed as a solution to the limited streaming capacity and null scalability of centralized systems. In a previous work, we proposed a fully distributed large-scale VoD architecture, called Double P-Tree, which has shown itself to be a good approach to the design of flexible and scalable DVoD systems. In this paper, we present relevant design aspects related to video mapping and traffic balancing in order to improve Double P-Tree architecture performance. Our simulation results demonstrate that these techniques yield a more efficient system and considerably increase its streaming capacity. The results also show the crucial importance of topology connectivity in improving multicasting performance in DVoD systems. Finally, a comparison among several DVoD architectures was performed using simulation, and the results show that the Double P-Tree architecture incorporating mapping and load balancing policies outperforms similar DVoD architectures.

This paper presents a communication system designed to allow efficient process migration in a cluster. The proposed system is generic enough to allow the migration of any kind of stream: socket, pipe, char devices. Communicating processes using IP or Unix sockets are transparently migrated with our mechanisms and they can still efficiently communicate after migration. The designed communication system is implemented as part of Kerrighed, a single system image operating system for a cluster based on Linux. Preliminary performance results are presented.

Smart Dust is comprised of a vast number of ultra-small fully autonomous computing and communication devices, with very restricted energy and computing capabilities, that co-operate to accomplish a large sensing task. Smart Dust can be very useful in practice i.e. in the local detection of a remote crucial event and the propagation of data reporting its realization to a control center. In this work, we have implemented and experimentally evaluated four protocols ( , and two variations of which we here introduce ) for local detection and propagation in smart dust networks, under new, more general and realistic modelling assumptions. We comparatively study, by using extensive experiments, their behavior highlighting their relative advantages and disadvantages. All protocols are very successful. In the setting we considered here, seems to be faster while the based protocols are more energy efficient.

We discuss a parallel implementation of an agent-based simulation. Our approach allows to adapt a sequential simulator for large-scale simulation on a cluster of workstations. We target discrete-time simulation models that capture the behavior of Web users and Web sites. Web users are connected with each other in a graph resembling the social network. Web sites are also connected in a similar graph. Users are stateful entities. At each time step, they exhibit certain behaviour such as visiting bookmarked sites, exchanging information about Web sites in the "word-of-mouth" style, and updating bookmarks. The real-world phenomena of emerged aggregated behavior of the Internet population is studied. The system distributes data among workstations, which allows large-scale simulations infeasible on a stand-alone computer. The model properties cause traffic between workstations proportional to partition sizes. Network latency is hidden by concurrent simulation of multiple users. The system is implemented in Mozart that provides multithreading, dataflow variables, component-based software development, and network-transparency. Currently we can simulate up to 10⁶ Web users on 10⁴ Web sites using a cluster of 16 computers, which takes few seconds per simulation step, and for a problem of the same size, parallel simulation offers speedups between 11 and 14.

Structured peer-to-peer (P2P) lookup services organize peers into a flat overlay network and offer distributed hash table (DHT) functionality. Data is associated with keys and each peer is responsible for a subset of the keys. In hierarchical DHTs, peers are organized into groups, and each group has its autonomous intra-group overlay network and lookup service. Groups are organized in a top-level overlay network. To find a peer that is responsible for a key, the top-level overlay first determines the group responsible for the key; the responsible group then uses its intra-group overlay to determine the specific peer that is responsible for the key. We provide a general framework for hierarchical DHTs with scalable overlay management. We specifically study a two-tier hierarchy that uses Chord for the top level. Our analysis shows that by using the most reliable peers in the top level, the hierarchical design significantly reduces the expected number of hops. We also present a method to construct hierarchical DHTs that map well to the Internet topology and achieve short intra-group communication delay. The results demonstrate the feasibility of locality-based peer groups, which allow P2P systems to take full advantage of the hierarchical design.

We consider the problem of designing optimal and efficient algorithms for solving tridiagonal linear systems with multiple right-hand side vectors on k-dimensional mesh and torus interconnection networks. We derive asymptotic upper and lower bounds for these solvers using odd-even cyclic reduction. We present various important bounds on execution time including general lower bounds which are independent of initial data assignment, and lower bounds based on classifying assignments via the proportion of initial data assigned amongst processors. Finally, different algorithms are designed in order to achieve running times that are within a small constant factor of the lower bounds provided.

Vehicle routing problems involve the navigation of one or more vehicles through a network of locations. Locations have associated handling times as well as time windows during which they are active. The arcs connecting locations have time costs associated with them. In this paper, we consider two different problems in single vehicle routing. The first is to find least time cost routes between all pairs of nodes in a network for navigating vehicles; we call this the all pairs routing problem. We show that there is an O(log³ n) time parallel algorithm using a polynomial number of processors for this problem on a CREW PRAM.

We next consider the problem in which a vehicle services all locations in a network. Here, locations can be passed through at any time but only serviced during their time window. The general problem is -complete under even fairly stringent restrictions but polynomial algorithms have been developed for some special cases. In particular, when the network is a line, there is no time cost in servicing a location, and all time windows are unbounded at either their lower or upper end, O(n²) algorithms have been developed. We show that under the same conditions, we can reduce this problem to the all pairs routing problem and therefore obtain an O(log³ n) time parallel algorithm on a CREW PRAM.

Given two genomic DNA sequences, the syntenic alignment problem is to compute an ordered list of subsequences for each sequence such that the corresponding subsequence pairs exhibit a high degree of similarity. Syntenic alignments are useful in comparing genomic DNA from related species and in identifying conserved genes. In this paper, we present a parallel algorithm for computing syntenic alignments that runs in time, where m and n are the respective lengths of the two genomic sequences, and p is the number of processors used. Our algorithm is time optimal with respect to the corresponding sequential algorithm and can use processors, where n is the length of the larger sequence. The space requirement of the algorithm is per processor. Using an implementation of this parallel algorithm, we report the alignment of a gene-rich region of human chromosome 12, namely 12p13 and its syntenic region in mouse chromosome 6 (both over 220,000 base pairs in length) in under 24 minutes on a 64-processor IBM xSeries cluster.

We propose a deadlock-free routing scheme in irregular networks using prefix routing. Prefix routing is a special type of routing with a compact routing table associated with each node (processor). Basically, each outgoing channel of a node is assigned a special label and an outgoing channel is selected if its label is a prefix of the label of the destination node. Node and channel labeling in an irregular network is done through constructing a spanning tree. The routing process follows a two-phase process of going up and then down along the spanning tree, with a possible cross channel (shortcut) between two branches of the tree between two phases. We show that the proposed routing scheme is deadlock- and livelock-free. We also compare prefix routing with the existing up*/down* routing which has been widely used in irregular networks. Possible extensions are also discussed.

This paper studies the problem of testing shared memory Java implementations to determine whether the memory behavior they provide is consistent. The complexity of the task is analyzed. The problem is defined as that of analyzing memory access traces. The study showed that the problem is NP-complete, both in the general case and in some particular cases in which the number of memory operations per thread, the number of write operations per variable, and the number of variables are restricted.