In this paper impact of the computing platform reliability on the performance of a stochastic optimization workflow is analyzed. A method to flexibly match the application computational demands with the computing platform variability is proposed. This goal is achieved by trading the optimization quality for the survival probability. This study originates in an optimization problem consisting in partitioning a maritime container terminal quay into berths so that the arriving vessel traffic waiting time is minimum. Thus, two rich domains – maritime logistics and HPC reliability – are connected in this work. Evaluation of a quay partition, as a part of the optimization process, is conducted in parallel on a Slurm cluster. Collected performance results revealed limitations of the workflow scalability and the reliability of the computing platform. This practical experience lead to the analysis of the performance and solvability limits on an unreliable platform. Though this study emerged in an effect of solving a port logistic problem in parallel, the analysis of this use-case is pertinent to other stochastic optimization applications run on computing platforms with lowered reliability such as data centers powered with green energy. The extent of the simulation and the number of optimization steps may be adjusted to the survival probability.

In this paper we consider filtering and processing large data streams in intelligent data acquisition systems.It is assumed that raw data arrives in discrete events from a single expensive sensor.Not all raw data, however, comprises records of interesting events and hence some part of the input must be filtered out.The intensity of filtering is an important design choice because it determines the complexity of filtering hardware and software and the amount of data that must be transferred to the following processing stages for further analysis.This, in turn, dictates needs for the following stages communication and computational capacity.In this paper we analyze the optimum intensity of filtering and its relationship with the capacity of the following processing stages.A set of generic filtering intensity, data transfer, and processing archetypes are modeled and evaluated.

We analyze time and energy performance of distributed computations in heterogeneous systems with hierarchical memory. Different levels of memory hierarchy have different time and energy efficiency. Core memory may be too small to hold whole load to be processed, while computations using external storage are expensive in time and energy. In order to avoid the costs of processing the load in the external memory, it is allowed that the load is distributed to the worker processors in multiple installments. A minimum energy solution is found by use of mixed integer linear programming under a limit on schedule length. Two types of fast heuristics with several variants are also examined. The trade-off between the criteria of processing time and energy is studied. Key features of optimum solutions are analyzed. It is shown that holding machines in a diverse set of energy modes and limited use of the out-of-core memory can be beneficial for the time and energy performance. The proposed scheduling algorithms are evaluated in the terms of solution quality and runtimes.

In this paper we analyze the impact of memory hierarchies on time-energy trade-off in parallel computations. Contemporary computing systems have deep memory hierarchies with significantly different speeds and power consumptions. This results in nonlinear phenomena in the processing time and energy usage emerging when the size of the computation is growing. In this paper the nonlinear dependence of the time and energy on the size of the solved problem is formalized and verified using measurements in practical computer systems. Then it is applied to formulate a problem of minimum time and minimum energy scheduling parallel processing of divisible loads. Divisible load theory is a scheduling and performance model of data-parallel applications. Mathematical programming is exploited to solve the scheduling problem. A trade-off between energy and schedule length is analyzed and again nonlinear relationships between these two criteria are observed. Further performance analysis reveals that energy consumption and schedule length are ruled by a complex interplay between the costs and speeds of on-core and out-of-core computations, communication delays, and activating new machines.

A CSS-sprite packing problem is considered in this article. CSS-sprite is a technique of combining many pictures of a web page into one image for the purpose of reducing network transfer time. The CSS-sprite packing problem is formulated here as an optimization challenge. The significance of geometric packing, image compression and communication performance is discussed. A mathematical model for constructing multiple sprites and optimization of load time is proposed. The impact of PNG-sprite aspect ratio on file size is studied experimentally. Benchmarking of real user web browsers communication performance covers latency, bandwidth, number of concurrent channels as well as speedup from parallel download. Existing software for building CSS-sprites is reviewed. A novel method, called Spritepack , is proposed and evaluated. Spritepack outperforms current software.

In this paper, we propose a new technique of presenting performance relationships in parallel processing. Performance of parallel processing is a hard matter with many counterintuitive phenomena. It is relatively easy to obtain some numerical indicators of the performance using various performance models. However, it is far more difficult to comprehend the nature of the analyzed problem. To facilitate understanding the performance relationships, we propose a new visualization technique based on the concept of isoefficiency. In this paper, isoefficiency is represented as a relation on points in the space of system parameters for which efficiency of parallel processing is equal. We visualize this relation on two-dimensional maps analogously to isobars and isotherms on weather maps. This concept is applied to depict the performance relationships in two standard performance laws: Amdahl's speedup law and Gustafson's speedup law. Then, we use isoefficiency maps to analyze the performance relationships in divisible load processing. Divisible load model conforms with data-parallel computations in an environment with communication delays. The results we obtain give interesting insights into relationships existing in parallel processing.

Abstract Divisible loads are parallel applications with fine granularity and negligible data dependencies. Such computations can be divided into parts of arbitrary sizes and processed independently in parallel. The load distribution process incurs considerable communication delays. To reduce processor waiting time during the computation initialization phase, the load is distributed in multiple small installments rather than in one big chunk. In this paper we analyze multi‐installment divisible load processing in heterogeneous distributed systems. Scheduling divisible loads in heterogeneous systems is hard because the sizes of the installments should be adjusted to the communication and computation capabilities of the system. We show that ignoring heterogeneity of the distributed system may result in arbitrarily bad solutions. Two algorithms are proposed to gear the load chunk sizes to different communication and computation speeds: an optimization branch‐and‐bound algorithm and a heuristic based on a genetic search method. The running times of both methods and the quality of the solutions are compared. Then, we use these algorithms to study the features of the multi‐installment divisible load scheduling problem. We demonstrate that it has both combinatorial and algebraic nature, and that optimum solutions are harder to find with the growing heterogeneity of the system. Copyright © 2007 John Wiley & Sons, Ltd.

In this paper, we analyze processing divisible loads in systems with a memory hierarchy. Divisible loads are computations that can be divided into parts of arbitrary sizes and these parts can be independently processed in a distributed system. The problem is to partition the load so that the total processing time, including communications and computations, is the shortest possible. Earlier works in the divisible load theory assumed distributed systems with a flat memory model. The dependence of the processing time on the size of the assigned load was assumed to be linear. A new mathematical model relaxing the above two assumptions is proposed in this article. We study distributed systems-which have both the hierarchical memory model and a piecewise linear dependence of the processing time on the size of the assigned load. Performance of such systems is modeled and evaluated. Finally, we compare the efficiency of distributed processing divisible loads in multiinstallment and out-of-core modes. Multiinstallment processing consists in sending multiple small chunks of the load to processors instead of a single chunk which needs external memory. It turns out that multiinstallment is an advantageous strategy for reasonably selected load chunks sizes.