Multiple sequence alignment is a fundamental and very computationally intensive task in molecular biology. MUSCLE, a new algorithm for creating multiple alignments of protein sequences, achieves a highest rank in accu...
详细信息
Particle tracing for streamline and path line generation is a common method of visualizing vector fields in scientific data, but it is difficult to parallelize efficiently because of demanding and widely varying compu...
详细信息
The H1N1 pandemic of 2009 and the ongoing Ebola outbreak in West Africa serve as a reminder of the social, economic and health burden of infectious diseases. The ongoing trends towards urbanization, global travel, cli...
详细信息
ISBN:
(纸本)9781479986484
The H1N1 pandemic of 2009 and the ongoing Ebola outbreak in West Africa serve as a reminder of the social, economic and health burden of infectious diseases. The ongoing trends towards urbanization, global travel, climate change and a generally older and immuno-compromised population continue to make epidemic planning and control challenging. Recent quantitative changes in high performance pervasive computing have created new opportunities for collecting, integrating, analyzing and accessing information related to large urban social systems, disease surveillance and global logistics and supply chains. The advances in network and information science that build on this new capability provide entirely new ways for reasoning and controlling epidemics. In this talk I will overview of the state of the art in computational networked epidemiology with an emphasis on computational thinking and on the development of high performance computing oriented decision-support environments for planning and response in the event of epidemics. I will describe how such systems can be used to support near real-time planning and response during the 2009 H1N1 swine flu and the recent Ebola Outbreak in West Africa. Computational challenges and directions for future research will be discussed.
Remotely sensed hyperspectral imaging is a technique that generates hundreds of spectral bands at different wavelength channels for the same area on the surface of the Earth. Computationally effective processing of th...
详细信息
ISBN:
(纸本)9781424456499
Remotely sensed hyperspectral imaging is a technique that generates hundreds of spectral bands at different wavelength channels for the same area on the surface of the Earth. Computationally effective processing of these image cubes can be greatly beneficial in many application domains, including environmental modeling, risk/hazard prevention and response, or defense/security. With the aim of providing an overview of recent developments and new trends in the design of parallel and distributed systems for hyperspectral image analysis, this paper discusses and inter-compares four different strategies for efficiently implementing a standard hyperspectral image processing chain: 1) commodity Beowulf-type clusters, 2) heterogeneous networks of workstations, 3) field programmable gate arrays (FPGAs), and 4) graphics processing units (GPUs). Combined, these parts deliver a snapshot of the state-of-the-art in those areas, and a thoughtful perspective on the potential and emerging challenges of adapting high performance computing systems to remote sensing problems.
Some applications, like round-based consensus algorithms, require all the nodes from a system to send a message to the same node (the leader) at the same time. In a Mobile Ad-Hoc Network (MANET), this situation is lik...
详细信息
McNaughton's theorem (1959) states that preemptions in scheduling arbitrary processing time jobs on identical parallel machines to minimize the total weighted completion time are redundant. Du, Leung and Young (19...
详细信息
General-purpose graphical processing units (GPGPUs) have become increasingly valuable platforms for accelerating parallel deep learning applications in recent years. GPUs are energy efficient if the software utilizes ...
详细信息
ISBN:
(纸本)9781665484855
General-purpose graphical processing units (GPGPUs) have become increasingly valuable platforms for accelerating parallel deep learning applications in recent years. GPUs are energy efficient if the software utilizes all usable resources;however, there are no hardware frameworks to adjust resources according to the application's requirements. In this paper, we address a robust methodology for optimizing the structure of GPGPUs using machine learning (ML) techniques. The candidate model allows the use of output measurements from the base GPU configuration to determine the optimal GPU resources between three classes (large, medium, and small). We infer from the experimental results that, while we can achieve substantial power-aware computing, they can be significant while running the program on its suitable GPU setup. The ML-classification approach is tested under different CUDA application workloads to validate its effectiveness as well as GPU energy consumption and instruction per cycle metrics.
Current processors exploit out-of-order execution and branch prediction to improve instruction level parallelism. When a branch prediction is wrong, processors flush the pipeline and squash all the speculative work. H...
详细信息
ISBN:
(纸本)0769523129
Current processors exploit out-of-order execution and branch prediction to improve instruction level parallelism. When a branch prediction is wrong, processors flush the pipeline and squash all the speculative work. However, control-flow independent instructions compute the same results when they re-enter the pipeline down the correct path. If these instructions are not squashed, branch misprediction penalty can significantly be reduced. In this paper we present a novel mechanism that detects control-flow independent instructions, executes them before the branch is resolved, and avoids their re-execution in the case of a branch misprediction. The mechanism can detect and exploit control-flow independence even for instructions that are far away from the corresponding branch and even out of the instruction window. Performance figures show that the proposed mechanism can exploit control-flow independence for nearly 50% of the mispredicted branches, which results in a performance improvement that ranges from 14% to 17,8% for realistic configurations of forthcoming microprocessors.
Real-time monitoring is increasingly becoming important in various scenes of large scale, multi-site distributed/parallel computing, e.g, understanding behavior of systems, scheduling resources, and debugging applicat...
详细信息
ISBN:
(纸本)9781424442379
Real-time monitoring is increasingly becoming important in various scenes of large scale, multi-site distributed/parallel computing, e.g, understanding behavior of systems, scheduling resources, and debugging applications. Dedicated networks on inter-site communications are rarely available for the monitoring purposes. Therefore, for real-time monitoring systems, reducing communication cost is important to handle a large number of nodes with limited network resources. We implemented a real-time Grid monitoring system called VGXP with techniques for low cost data gathering. It tries to send only diffs to recent data, and adapts to the requested data freshness and tolerable errors to minimize required communication. We evaluate monitoring overheads of the proposed method on a distributed environment consisting of 8-sites with 500 nodes. In a realistic setting where the sampling interval is set to 0.5 seconds and the tolerable error to 2%, the CPU usage of the server to gather data from all nodes was 0.2% and the transfer rate was less than 5kbps. The transfer rate did not exceed 50kbps even if we gather a detailed per-process statistics.
Disk power consumption is becoming an increasingly important issue in high-end servers that execute large-scale data-intensive applications. In particular, array-based scientific codes can spend a significant portion ...
详细信息
ISBN:
(纸本)0769523129
Disk power consumption is becoming an increasingly important issue in high-end servers that execute large-scale data-intensive applications. In particular, array-based scientific codes can spend a significant portion of their power budget on the disk subsystem. Observing this, the prior research proposed several strategies, such as spining down to low-power modes or adjusting the speed of the disk in lower RPM, to reduce power consumption on the disk subsystem. A common characteristic of most of these techniques is that they are reactive, in the sense that they make their decisions based on the disk access patterns observed during execution. While such techniques are certainly useful and the published studies reveal that they can be very effective in some cases, one can conceivably achieve better results by adopting a proactive scheme. Focusing on array-intensive scientific applications, this paper makes two important contributions. First, it presents a compiler-driven proactive approach to disk power management. In this approach, the compiler analyzes the application code and extracts the disk access pattern. It then uses this information to insert explicit disk power management calls in the appropriate places in the code. It also preactivates a disk (placed into the low-power mode) before it is actually needed to eliminate the potential performance impact of disk power management. The second contribution of this paper is a code transformation approach that can be used to increase the savings coming from a disk power management scheme (whether reactive or proactive). Our experimental results with several scientific application codes show that both the proactive disk power management approach and the disk layout aware code transformations are beneficial from both power consumption and execution time perspectives.
暂无评论