Search

Big-data application processing is increasingly geo-distributed, a paradigm shift from the traditional cluster-based processing frameworks. As the communication time for data movement across geo-distributed data centers is not a design criterion for traditional cluster-based processing approaches, there are research gaps in the algorithms used for staging and scheduling big-data applications for geo-distributed clusters. We address these gaps by proposing Harmony, an approach consisting of both staging and scheduling strategies to minimize an application’s total execution time. The staging strategy of Harmony exploits the intra-stage parallelism by having concurrent operators within a stage in contrast to the traditional Apache spark which uses fine-grained stages, thus reducing the computation time within each stage. Secondly, the scheduling strategy of Harmony reduces data transfers between geo-distributed data centers by exploiting data locality and thus reducing communication time and total execution time. The proposed approach Harmony achieves a speedup of two times with respect to geo-distributed Apache Spark. In addition, Harmony achieves a speedup of 1.6 times and 2.1 times when compared with the state-of-the-art framework Iridium for geo-distributed analytics over five locations with uniform and non-uniform network link bandwidths respectively.

While edge computing is gaining traction, organizations operating in geographically distributed locations are still using cloud computing to collect and post-process data. In this context, it is useful to analyze the performance trade-offs of cloud-only, edge-only and hybrid edge-cloud processing. To facilitate this analysis, we provide an analytic model validated by measurements on representative edge and cloud platforms. Our model is easy to apply even without performing measurements on the target edge hardware, as long as useful performance specifications are available. Our measurement-driven analysis reveals a diverse performance landscape where there is no clear winner among cloud-only, edge-only and hybrid processing. However, application characteristics and edge-cloud transfer bandwidth are the key factors affecting performance.

Recent advancements in big data processing and machine learning, among others, increase the resource demand for running applications with larger problem sizes. Elastic cloud computing resources with pay-per-use pricing offers new opportunities where large application execution is constrained only by the cost budget. Given a cost budget and a time deadline, this paper introduces a measurement-driven analytical modeling approach to determine the largest Pareto-optimal problem size and its corresponding cloud configuration for execution. We evaluate our approach with a set of representative applications that exhibit a range of resource demand growth patterns on Amazon AWS cloud. We show the existence of cost-time-size Pareto-frontier with multiple sweet spots meeting user constraints. To characterize the cost-performance of cloud resources, we use Performance Cost Ratio (PCR) metric. We extend Gustafson's fixed-time scaling in the context of cloud, and, investigate fixed-cost-time scaling of applications and show that using resources with higher PCR yields better cost-time performance. We discuss a number of useful insights on the trade-off between the execution time and the largest Pareto-optimal problem size, and, show that time deadline could be tightened for a proportionately much smaller reduction of problem size.

Clouds offer great flexibility for scaling applications due to the wide spectrum of resources with different cost-performance, inherent resource elasticity and pay-peruse charging. However, determining cost-time-efficient cloud configurations to execute a given application in the large resource configuration space remains a key challenge. The growing importance of elastic applications for which the accuracy is a function of resource consumption introduces new opportunities to exploit resource elasticity on clouds. In this paper, we introduce CELIA, a measurement-driven analytical modeling approach to determine cost-time-optimal cloud resource configurations to execute a given elastic application with a time deadline and a cost budget. We evaluate CELIA with three representative elastic applications on more than ten million configurations consisting of Amazon EC2 resource types with different cost-performance. Using CELIA, we show that multiple cost-time Pareto-optimal configurations exist among feasible cloud configurations that execute an elastic application within a time deadline and cost budget. These Pareto-optimal configurations exhibit up to 30% cost savings for an elastic application representing n-body simulation. We investigate the impact of fixed-time scaling on the cost of executing elastic applications on cloud. We show that cost gradient with respect to resource demand is smaller when cloud resources with better cost-performance are used. Furthermore, we show that the relative increase in cost is always smaller compared to the relative reduction of execution time deadline. For example, tightening the execution time deadline by two-thirds incurs only 40% increase in cost for the n-body simulation application.

Motivated by the increasing importance of edge computing and improvements in heterogeneous low-power, wimpy systems, we present a novel time-energy-cost analysis of wimpy edge computing in comparison with traditional brawny cloud computing. For this analysis, we use a brawny heterogeneous Amazon EC2 cloud instance with GPU and two wimpy heterogeneous systems represented by Jetson TK1 and Jetson TX1. As the paradigm shift to edge computing is due to the challenges of Big Data processing, we select six representative MapReduce applications with applicability in IoT edge computing. Using our time-energy-cost analysis of both wimpy edge compute nodes and cloud computing systems with GPUs, we present several key insights. Firstly, we advocate the usage of heterogeneous systems with GPU on both edge and cloud since they provide time-energy savings of up to 70% for compute-intensive applications. Secondly, we establish an equivalence ratio between one brawny cloud instance and multiple wimpy edge nodes that achieve the same or better time performance. Based on this equivalence ratio, we show that using wimpy systems as edge computing devices saves cost compared to using traditional cloud computing. Lastly, counter-to-intuition, our analysis shows that the latest Jetson TX1 system exhibits worse time-cost performance compared to the older Jetson TK1 system. This result stems from lower operating core clock frequency and lower instructions-per-cycle of Jetson TX1's GPU on some compute-intensive applications.

Quantum Zeno blockade offers a distinct approach to signal manipulation and logic operations in a counterintuitive “interaction-free” implementation. Previous experimental studies have used nonlinear waveguides and optical cavities, both in bulk-optics settings. We report the observation of quantum Zeno blockade in a chip-integrable platform, whose results point to all-optical information processing on a single photon level.

Many Integrated Core (MIC) architecture systems are becoming increasingly popular for HPC applications as they have the dual-advantage of accelerating vector processing and a general-purpose programming model. One of the key challenges for energy-efficient execution on MIC architecture systems is to determine time and energy-efficient configurations among a large system configuration space. Given a parallel program with an execution time deadline and an energy budget, we propose a measurement-based analytical modeling approach to determine these system configurations. In contrast to current approaches, we model both inter-and intra-core resource overlaps, memory contention among threads within a core and memory contention across multiple cores. The model is validated against direct measurement using six representative HPC applications on Intel Xeon Phi system. We show that a Pareto frontier consisting of optimal configurations exist for a parallel program executing on MIC architecture systems. To further understand the Pareto frontier, we use the performance-to-power ratio metric (PPR), that quantifies the amount of useful computations performed per unit energy during the execution. The proposed approach can be used to determine what thread affinity is suitable for offloading execution to accelerators such as Xeon Phi and save energy.

Energy efficiency is an area of growing importance in datacenters. While the energy proportionality wall poses a challenge to further improve the dynamic power range of servers, heterogeneous systems offer a new opportunity to achieve higher energy efficiency by improving the match between application workload demands on the large heterogeneous system configuration space. This paper proposes a model-driven energy proportionality analysis of heterogeneous clusters consisting of server nodes with different performance-to-power ratio (PPR). Our analysis shows that inter-node heterogeneity has a positive effect of scaling the energy proportionality wall by exposing configurations with sub-linear energy proportionality. Secondly, analysis of these sub-linear configurations on the 95th percentile response time shows that heterogeneity is beneficial in workloads where the PPR of wimpy nodes is higher than brawny nodes.

An extended GIHS (EGIHS) fusion approach is presented for merging the panchromatic (PAN) image and the multi-spectral (MS) image with model-based optimization (MBO). The pan-sharpened MS image is posed as the optimization solution of the proposed functional, which consists of two energy terms. The first energy term injects the details of the PAN image into the MS image with the generalized IHS (GIHS) fusion model and the second energy term preserves the spectral contents of the MS image by enforcing the low-pass version of the pan-sharpened image should be close to the original up-sampled MS image, in which the low-pass filter is developed based on the modulation transfer function (MTF) of the different MS band. The pan-sharpened MS images are decoupled in the proposed method, which allows an efficient implementation. Experiments on IKONOS and QuickBird datasets demonstrate that the proposed approach outperforms the state-of-the-art fusion models including IHS-based GIHS, GIHSA, IHS-BT, MBO-based Adaptive IHS and MRA-based AWLP.

“The whole is greater than the sum of its parts” is attributed to the Greek philosopher Aristotle. It is well known that entity interactions in a system give rise to new behavior of the whole or emergent properties that may not be reducible to the behavior of the individual entities. As system complexity grows in terms of size and scale, advancing our understanding of emergent properties is becoming important and challenging. This keynote presents our work in addressing some of these challenges including formalizing weak emergence and the modelling of known emergence. Among the different perspectives of emergence from philosophy to natural sciences, we present a computer science perspective in formalizing weak emergence [3]. This approach is based on the understanding of weak emergence in scientific discussion where the truth in high-level phenomenon is unexpected given the principles governing low-level domain but is deducible via simulation. Without prior knowledge of emergence, a system is modelled as a multi-agent system and cooperative-array grammar to obtain the set of emergent property states. Secondly, we discuss different methods for quantifying the strength of these emergent property states as a mean for managing the large state-space and for emergent discovery [2]. Lastly, we present an approach to model emergence when there is prior knowledge of emergence. Using Twitter social network as an example, we present two new models: a Twitter users posting behavior model and a model for characterizing Twitter topic emergence and its intensity [1].

The saying The whole is greater than the sum of its parts is often attributed to the Greek philosopher Aristotle. In a system, entity interactions give rise to new behavior of the whole or emergent behavior that cannot be reducible to the behavior of the individual entities. As systems grow in size and scale, advancing our understanding of emergent properties is both important and challenging.This keynote presents our work in addressing some of these challenges including formalizing weak emergence, quantifying the strength of emergence and modeling known emergence. Firstly, we present a computer science perspective in formalizing weak emergence [3].This approach using cooperative-array grammar can be applied without prior knowledge of emergence but suffers from state-space explosion. Secondly, we propose methods for quantifying the strength of emergent properties with the aim of reducing the state-space and for classifying emergent property states in emergent discovery [2]. Lastly, we show that with prior knowledge of emergence, state-space is significantly reduced.Using Twitter social network as an example, we present two models: posting behavior model of Twitter users and a model for characterizing Twitter topic emergence and its intensity [1].

Traditional datacenter systems advocate the use of high-performance hardware, resulting in increased power consumption and cooling costs. With increasing availability of systems having diverse performance-to-power ratios, we analyze the energy efficiency of mixing high-performance and low-power nodes in a cluster. Using a model-driven analysis, we predict the heterogeneous mix of nodes that is the most energy-efficient while maintaining a given deadline. Considering service demands of the workloads on cores, memory and I/O devices, we derive Pareto-optimal configurations by matching the execution rate of different nodes. Our mix and match approach determines heterogeneous configurations that exhibit a "sweet region", where energy usage reduces linearly as the deadline is relaxed. Our analysis shows that mixing high-performance and low-power nodes is more energy-efficient than homogeneous datacenter clusters.

The Internet is a pervasive medium for content distribution. It is estimated that in 2015, it will take us five years to view all video crossing IP networks each second. When new content is made available, file distribution systems often have to cope with a sudden surge in the number of users, called flash crowd, without loss of download performance. Flash crowds can have serious consequences on business revenue. For example, Amazon estimates that its sales decrease by 1% for every 100ms delay due to flash crowds, and Google reports that half a second increase in waiting time results in a 20% decrease in traffic. This keynote discusses a new approach in modeling the performance of peer-assisted file distribution systems with flash crowds [1, 3]. Peerassisted file distribution systems extend client-server systems [2, 4]. The peers download a file as a client but at the same time act like servers by sharing the file across peers, thus improving the overall system download performance. This keynote is divided into three main parts. First, we review three main approaches in analysis flash crowd performance, namely, measurement, simulation and analytic models. Secondly, based on insights drawn from extensive measurement studies, we propose a general analytical model for understanding flash crowd performance [1, 3]. We show that the utilization of available peer bandwidth over the duration of flash crowd can be characterized by three distinct phases called startup, maximum utilization and end-game. We discuss the applications of our model by users, service providers and protocols designers of peer-assisted file distribution systems. In conclusion, we highlight a number of challenges and opportunities in modeling flash crowd performance in web-based and mobile applications.

Traditionally, imperative programming uses a series of state-based operands to model control-flow and, as a result, suffers from the well-known von Neumann bottleneck. In contrast, dataflow programs are driven only by the availability of instruction operands. However, the lack of mainstream dataflow hardware hinders direct dataflow instruction execution. On the other hand, direct execution of dataflow programs on von Neumann machines incurs a high performance cost. In this paper, we present preliminary results on the direct execution of dataflow programs on multicore systems through emulation of tagged-tokens mechanism. Compared with direct translation of SISAL programs to C code, we achieved a speedup of 44 for CPU-intensive applications and 22 for memory-bounded applications on a 48-cores AMD NUMA system.

With the rapid growth of urbanization, cities are stretching and expanding themselves, naturally followed by the risks and dangers in city operation, which poses an immense potential threat to its stable, proper and orderly operation. This paper will probe into the major issues confronting the municipalities, enterprises and the public. On the basis of the principle of """"Precaution goes first"""", we attempt to come up with a set of countermeasures in order to ensure safe operation of a city and to perfect the measures against risks and danger in the city under the normal circumstances with the aim to enhance anti-risk capacity.

Component-based models have been shown to exhibit emergent properties but despite a plethora of definitions and methods to identify emergence, practical semantic validation approaches remain a key challenge. This paper proposes an objective-based approach for semantic validation of emergence in component-based simulation models. In contrast to current methods, our approach describes model components in terms of what they achieve rather than how they achieve it, and we exploit reconstruct ability analysis in the validation of emergence. This has the advantage of reducing the number of attributes used to describe each component, and thus facilitates the application of more rigorous mathematical formalisms for emergence validation. As an example, we detail how this methodology is integrated into the life-cycle of our component-based model development framework.

With cloud computing, the long-envisioned dream of computing as utility is achieved. Many of the current standalone cloud providers offer resources and services using pay-per-use fixed pricing. Spot pricing, recently introduced in Amazon EC2, is more efficient by setting the resource price dynamically, based on demand. However, this pricing scheme, similar to the uniform price auction, is truthful only when supply can be adjusted, such as the case of a standalone cloud provider. In a federated cloud, where resources from multiple cloud providers are integrated to increase elasticity and reliability, rational users can have a greater impact. In this paper, we evaluate the impact of rationality in a federated cloud, by comparing the consumer welfare achieved by spot pricing, currently used in Amazon EC2, and a strategy-proof pricing scheme. We consider different EC2 regions as providers in a federated cloud, and use traces of spot prices to determine the consumer requests and the user welfare. Our results show that spot pricing is not suitable in a federated cloud, where rational users can increase their welfare by being untruthful.

Multicore systems are increasingly adopted across many application domains. Consequently, understanding their performance is becoming an important issue for a growing number of users. However, performance analysis of parallel programs on multicore systems is still challenging, especially for large programs or applications developed in multiple programming languages. This paper proposes an analytical modelling approach for studying the parallelism and energy performance of shared-memory programs on multicore systems. The proposed model derives the speedup and speedup loss from data dependency and memory overhead in traditional UMA and NUMA multicore systems, and emerging platforms such as ARM multicores. Using only widely available inputs derived from the trace of the operating system run-queue and hardware events counters, the proposed model achieves high practicality and generality across many types of shared-memory programs running on different multicore platforms. Applications of the model include understanding achieved speedup and parallelism loss, and prediction of optimal core and memory configuration, where the optimality criteria is minimum execution time, minimum energy usage or a trade-off between these two.

Peer-to-peer file-sharing protocols, such as BitTorrent, have been widely used to improve the performance and scalability of file distribution systems. In this paper, we study the performance of peer-assisted file distribution systems with heterogeneous peers. Based on a measurement study of BitTorrent on PlanetLab, we made two key observations: (i) there is a fixed pattern in the utilization of the available bandwidth over the course of a download, and (ii) peers enjoy an amount of service that is commensurate with their contribution. Building on these insights, we developed an analytical model to estimate the download time for each class of peers in a well-provisioned peer-assisted file distribution system based on BitTorrent. Our model accurately predicts the download time and achieves an average error rate of 16.5% for heterogeneous swarms up to 150 nodes in size. We demonstrate how it can be used to estimate the server capacity for achieving a specific quality of service in a large heterogeneous swarm.

A 3D natural seepage model of a hydropower station was established based on the engineering hydrogeology, the sequential quadratic programming was used to invert the natural seepage field and permeability parameters of the hydropower station. The inversion permeability parameters and reasonable intercepting boundaries were obtained, and the boreholes, whose head difference between the inversion and in-situ is below 20m, amount to the 80% of the total. This indicates that the natural seepage model can reflect the main engineering geological and groundwater distribution features of the hydropower station, the obtained permeability parameters and boundaries are reasonable.

Memory contention is an important performance issue in current multicore architectures. In this paper, we focus on understanding how off-chip memory contention affects the performance of parallel applications. Using measurements conducted on state-of-the-art multicore systems, we observed that off-chip memory traffic is not always bursty, as it was previously reported in literature. Burstiness depends on the problem size. Small problem sizes lead to bursty memory traffic, and generate small off-chip contention. In contrast, when large program sizes cause memory contention, the memory traffic is non-bursty. Based on these observations, we propose an analytical model that relates the growth of memory contention to the number of active cores and to the problem size, for both uniform (UMA) and non-uniform memory access (NUMA) systems. Our model differs from measurements on average by less than 14\%. Contention for off-chip memory grows exponentially with the number of active cores, but adding additional memory controllers reduces the memory contention. For programs such as the penta diagonal solver SP from NPB benchmark, with a large matrix of $162^3$ elements (input size C), our analysis shows that memory contention increases the total number of processor cycles to execute the program by more than ten times on a machine with 24 cores.

Validation of semantic composability is a non-trivial problem and a key step in component-based modeling and simulation. Recent work in semantic composability validation promise to reduce verification, validation, and accreditation efforts. However, the underlying cost of current validation approaches can undermine the promised benefits, and the trade-off between validation accuracy and validation cost is not well understood. In this paper we present, to the best of our knowledge, the first quantitative study on the cost of validating semantic composability. Firstly, validation approaches are categorized into techniques that validate general model properties, and techniques that validate model execution. Secondly, we focus on two key factors that influence validation cost, namely, simulation problem characteristics and the validation approach adopted. Our study covers four representative validation approaches, CD++ DEVS, Z-based DEVS, Petty and Weisel formal validation, and deny-validity, and for simplicity, we use computation time as a measure of validation cost. Based on a queueing network model with 1,000 components, the cost of validating general model properties accounts for 55% of total validation time, with the remaining 45% incurred by model execution validation. With respect to composition structure, a 10% increase in the fork and join component interconnections increases validation cost by more than half. In model execution validation, the time-based deny-validity approach is seven times more expensive than the timeless Petty and Weisel formalism.