Revisiting workflow execution in HPC: a data-flow approach

The skyline is an optimization operator widely used for multi-criteria decision making. It allows minimizing an n-dimensional dataset into its smallest subset. In this work we present SkyFlow, the first heterogeneous CPU+GPU graph-based engine for skyline computation on a stream of data queries. Two data flow approaches, Coarse-grained and Fine-grained, have been proposed for different streaming scenarios. Coarse-grained aims to keep in parallel the computation of two queries using a hybrid solution with two state-of-the-art skyline algorithms: one optimized for CPU and another for GPU. We also propose a model to estimate at runtime the computation time of any arriving data query. This estimation is used by a heuristic to schedule the data query on the device queue in which it will finish earlier. On the other hand, Fine-grained splits one query computation between CPU and GPU. An experimental evaluation using as target architecture a heterogeneous system comprised of a multicore CPU and an integrated GPU for different streaming scenarios and datasets, reveals that our heterogeneous CPU+GPU approaches always outperform previous only-CPU and only-GPU state-of-the-art implementations up to 6.86×and 5.19×, respectively, and they fall below 6% of ideal peak performance at most. We also evaluate Coarse-grained vs Fine-Grained finding that each approach is better suited to different streaming scenarios.

Multiple deep neural networks (DNNs) are increasingly used in real-world intelligent applications, such as intelligent robotics and autonomous vehicles to collectively complete complicated tasks running on edge devices. Because each layer of the subtasks prefers a distinct dataflow due to the heterogeneity in shape and scale of the network layers, a variable dataflow approach on the DNN accelerators is urgently required. On DNN accelerators that enable multiple dataflows, however, we detect a dimension mismatch between parallel processing under the dataflow approach and linear data memory arrangement. When multiple DNN tasks share partial features or weights, the issue is further exacerbated. During processing, this mismatch causes a sluggish data supply from both off-chip and on-chip memory. Consequently, the overall throughput, performance, and energy efficiency suffer since DNN models are sensitive to data density. In this work, we reveal the mechanism behind this data dimension mismatch and present a series of metrics that quantify the influence on system performance. On this foundation, we offer a framework that tracks the data tensor dimension conversion and employs a flexible data arrangement over multi-DNN computation to adapt to dataflow variability. An accelerator architecture named data arrangement multi-DNN accelerator (DARMA) that features a data arrangement and distribution circuit and hierarchical memory for data dimension conversion is also presented. Since the mismatch is mitigated, the suggested accelerator outperforms current accelerators in terms of bandwidth and processing unit utilization. Through tests on VR/AR, MLperf, and other multitask applications, the evaluation results show that the proposed architecture provides both energy-efficiency and throughput improvements.

Sleptsov nets are applied as a uniform language to specify models of unconventional computations and artificial intelligence systems. A technique for specification of neural networks, including multidimensional and multilayer networks of deep learning approach, using Sleptsov nets, is shown; the ways of specifying basic activation functions by Sleptsov net are discussed, the threshold and sigmoid functions implemented. A methodology of training neural networks is presented with the loss function minimisation, based on a run of a pair of interacting Sleptsov nets, the first net implementing the neural network based on data flow approach, while the second net solves the optimisation task by adjusting the weights of the first net by the gradient descend method. The optimising net uses the earlier developed technology of programming in Sleptsov nets with reverse control flow and the subnet call technique. Real numbers and arrays are represented as markings of a single place of a Sleptsov net. Hyperperformance is achieved because of the possibility of implementing mass parallel computations.

ABSTRACT The use of Smart Homes has grown considerably in the past decade. Enabling end-users to develop rules to program their homes and devices is very important to empower them. Several studies have analysed trigger-action programming tools, primarily using form-based and data-flow approaches for programming interfaces. This study evaluated the usability of a block-based tool for end-user development of rules to control smart homes and compared the difficulties encountered by non-programmers and programmers. Evaluations involved 10 programmers and 10 non-programmers in Brazil. A thematic analysis of 247 problem instances (80 from programmers and 167 from non-programmers) yielded the following themes, with problems related to condition blocks, action blocks, states and actions, time-related tasks, block configuration and personalisation, information architecture, programming logic, conceptual model of smart homes, simulator and debugging, help and technical problems. Despite most non-programmers being able to experiment with blocks, their task completion rates were significantly lower than programmers. The analysis showed aspects where block-based programming can enhance the use for non-programmers. They also confirmed interaction aspects revealed by previous studies using form-based and data-flow approaches that also occur with block-based programming to design smart home rules, with important contributions to improve end-user development tools for smart homes.

Deep neural networks (DNNs) come with many forms, such as convolutional neural networks, multilayer perceptron and recurrent neural networks, to meet diverse needs of machine learning applications. However, existing DNN accelerator designs, when used to execute multiple neural networks, suffer from underutilization of processing elements, heavy feature map traffic, and large area overhead. In this paper, we propose a novel approach, Polymorphic Accelerators, to address the flexibility issue fundamentally. We introduce the abstraction of logical accelerators to decouple the fixed mapping with physical resources. Three procedures are proposed that work collaboratively to reconfigure the accelerator for the current network that is being executed and to enable cross-layer data reuse among logical accelerators. Evaluation results show that the proposed approach achieves significant improvement in data reuse, inference latency and performance, e.g., 1.52x and 1.63x increase in throughput compared with state-of-the-art flexible dataflow approach and resource partitioning approach, respectively. This demonstrates the effectiveness and promise of polymorphic accelerator architecture.

As the advent of the big data era, huge-scale data continuously appears in various fields of science, commerce, industry and society. More algorithms/methods/approaches are urgently required to learn huge-scale data collected from different applications/backgrounds. Therefore, the Pseudo Data Flow (Pseudo-DF) approach with ensemble ReOS-ELMs is proposed in this paper. The Pseudo-DF approach randomly divides a huge-scale data set into K (K>1) non-overlapping data chucks, and a Pseudo-DF is constructed by these data chucks. The computation of a huge-scale data is changed into that of a Pseudo-DF with smaller-scale chucks, the computational burden will be much reduced. Then, the ensemble Regularized OS-ELMs (ReOS-ELMs) based on Different random Hidden-node Parameters (DiffHPs) is presented to learn a Pseudo-DF, which is a recursive leaning algorithm possessing the advantages of low computational burden, high accuracy, well generalization and stability, and strong robustness. Lastly, experiments are performed to validate the effectiveness of the proposed approach.

The Partitioned Global Address Space (PGAS) programming model is one of the most relevant proposals to improve the ability of developers to exploit distributed memory systems. However, despite its important advantages with respect to the traditional message-passing paradigm, PGAS has not been yet widely adopted. We think that PGAS libraries are more promising than languages because they avoid the requirement to (re)write the applications using them, with the implied uncertainties related to portability and interoperability with the vast amount of APIs and libraries that exist for widespread languages. Nevertheless, the need to embed these libraries within a host language can limit their expressiveness and very useful features can be missing. This paper contributes to the advance of PGAS by enabling the simple development of arbitrarily complex task-parallel codes following a dataflow approach on top of the PGAS UPC++ library, implemented in C++. In addition, our proposal, called UPC++ DepSpawn, relies on an optimized multithreaded runtime that provides very competitive performance, as our experimental evaluation shows.

IoT devices are diverse in their characteristics and made by many vendors, hence the inter-operation among them is difficult. Especially, end users can’t make their own programs by do-it-yourselves. IFTTT and Zapier platforms are designed to help end users to make them inter-operable easily and prevail in these days. Their approach is categorized into a Trigger-Action-Programming, in which trigger conditions and actions are already made by professional programmers of several IoT vendors and end users composite them into their own applications easily. But, their drawback is that the composition can be made at once in the first level, hence end users can’t make more complicated applications. Our approach is based on a dataflow programming paradigm which resembles the TAP in that the internal actions are triggered when all the inputs of a node are prepared. In our approach, a composition of some atomic nodes becomes another atomic node, so the composition would continue iteratively. This feature is so generous that several visual programming languages like LabView are relied on the approach for various fields. We propose the overall architecture of our system and explain them. We also present Internet of Things examples of our approach, which shows that atomic dataflow objects can be associated to produce composite dataflow objects. And they are also composited to make more complex applications iteratively. We compare IFTTT, Zapier, and our approach qualitatively and show that end users can make more diverse and flexible applications in our approach.

We present an approach to enable end-users to graphically compose their own applications directly on their mobile phone, mainly integrating the functionalities available on the device and those provided by pervasive and Internet services. To this aim, we propose a methodology and a graphical notation enabling the user to compose mobile applications, named MicroApps: the user creates an application following an incremental and iterative development process; he composes icons representing (pervasive) services mainly by touch-based selection and following a data-flow approach. He is not in charge of the creation of the user interface, which is automatically generated. The methodology enables the end-user to develop applications and/or compose services on the smartphone, so paving the way towards new scenarios where smartphones replace and overtake the Personal Computer, given their native possibility of wide connectivity, when augmented by features for interaction with remote systems and sensors. The methodology has been evaluated through an empirical analysis that revealed that in spite of the reduced size of the screen the use of the MicroApp Generator tool improves the effectiveness in terms of time and editing errors with respect to the use of MIT App Inventor [1] .

This paper describes the vision behind and the mission of the Maxeler Application Gallery (AppGallery.Maxeler.com) project. First, it concentrates on the essence and performance advantages of the Maxeler dataflow approach. Second, it reviews the support technologies that enable the dataflow approach to achieve its maximum. Third, selected examples of the Maxeler Application Gallery are presented; these examples are treated as the final achievement made possible when all the support technologies are put to work together (internal infrastructure of the AppGallery.Maxeler.com is given in a follow-up paper). As last, the possible impact of the Application Gallery is presented and the major conclusions are drawn.

Computational architectures for sonar array processing in autonomous rovers

SummaryComputer simulations may ingest and generate high numbers of raw data files. Most of these files follow a de facto standard format established by the application domain, for example, Flexible Image Transport System for astronomy. Although these formats are supported by a variety of programming languages, libraries, and programs, analyzing thousands or millions of files requires developing specific programs. Database management systems (DBMS) are not suited for this, because they require loading the raw data and structuring it, which becomes heavy at large scale. Systems like NoDB, RAW, and FastBit have been proposed to index and query raw data files without the overhead of using a database management system. However, these solutions are focused on analyzing one single large file instead of several related files. In this case, when related files are produced and required for analysis, the relationship among elements within file contents must be managed manually, with specific programs to access raw data. Thus, this data management may be time‐consuming and error‐prone. When computer simulations are managed by a scientific workflow management system (SWfMS), they can take advantage of provenance data to relate and analyze raw data files produced during workflow execution. However, SWfMS registers provenance at a coarse grain, with limited analysis on elements from raw data files. When the SWfMS is dataflow‐aware, it can register provenance data and the relationships among elements of raw data files altogether in a database, which is useful to access the contents of a large number of files. In this paper, we propose a dataflow approach for analyzing element data from several related raw data files. Our approach is complementary to the existing single raw data file analysis approaches. We use the Montage workflow from astronomy and a workflow from Oil and Gas domain as data‐intensive case studies. Our experimental results for the Montage workflow explore different types of raw data flows like showing all linear transformations involved in projection simulation programs, considering specific mosaic elements from input repositories. The cost for raw data extraction is approximately 3.7% of the total application execution time. Copyright © 2015 John Wiley & Sons, Ltd.

Code based (“white box”) approach to testing can be divided into two main types: control flow coverage and data flow coverage methods. Dataflow testing was introduced for structural programming languages and later adopted for object languages. Among many tools supporting code based testing of object programs, only JaBUTi and DFC (Data Flow Coverage) support dataflow testing of Java programs. DFC is a tool implemented at the Institute of Computer Science Warsaw University of Technology as an Eclipse plug-in. The objective of this paper is to present dataflow coverage testing of Java programs supported by DFC. DFC finds all definition-uses pairs in tested unit and provides also the definition-uses graph for methods. After the execution of test information which def-uses pairs were covered is shown. An example of data flow testing of Java program is also presented.

Cluster-to-cluster data transfer with data compression over wide-area networks

A meta-model for dataflow-based rules in smart environments: Evaluating user comprehension and performance

Hot Path Dataflow Approach to Software Watermarking

Abstract This paper illustrates the theoretical basis of an approach to apply data flow testing techniques to abstract state machines (ASMs). In particular, we focus on multi-agent ASMs extended with theseqconstruct for turbo ASMs. We explain why traditional data flow analysis can not simply be applied to ASMs: data flow coverage criteria are strictly based on the mapping between a program and its flow graph whereas in this context we are interested in tracing the flow of data between states in ASM runs as opposed to between nodes in a program’s flow graph. We revise the classical concepts in data flow analysis taking into account the specific, parallel nature of ASMs, and define them on two levels: the syntactic (rule) level, and the computational (run) level. In particular, we analyze the role played by different types of terms in ASMs and deal with the problem of terms that are monitored by a given agent but controlled by another one, terms that are shared between several agents, and derived terms. We also discuss what consequences the use of the turbo ASM constructseqhas on our analysis and revise the approach accordingly. Finally, we specify a family of ad hoc data flow coverage criteria for this class of ASMs and introduce a model checking-based approach to generate automatically test cases satisfying a given set of coverage criteria from ASM models.

This paper presents the OpenDF framework and recalls that dataflow programming was once invented to address the problem of parallel computing. We discuss the problems with an imperative style, von Neumann programs, and present what we believe are the advantages of using a dataflow programming model. The CAL actor language is briefly presented and its role in the ISO/MPEG standard is discussed. The Dataflow Interchange Format (DIF) and related tools can be used for analysis of actors and networks, demonstrating the advantages of a dataflow approach. Finally, an overview of a case study implementing an MPEG- 4 decoder is given.

A teaching-learning process formalized through the IMS-Learning Design specification (IMS-LD) comprises a sequence of learning activities (learning flow) as well as a sequence of artifacts between tools or services (data flow) used to support the learning activities. According to the literature, the collaborative learning flow specification has been successfully achieved; however the automation of the collaborative data flow is still an open issue in IMS-LD. Nevertheless, no case studies have been reported in the literature in order to show with real data why the automation of the data flow is an important issue in Learning Design (LD). In this paper an authentic case study which is significative and relevant to the problem is analyzed. Supported with real data, several findings related to the data flow problem in collaborative learning emerged: data management is error-prone for the users; data flow specification is error-prone for the course designer; users suffer an additional cognitive load during the data management; the course designer suffers an additional cognitive load during the data flow specification; and the need to include instance-level data flow specification within the learning design has also been identified. Furthermore, these findings also help us to understand the relevance of the problem: a data flow approach which is error-prone for both the users and the course designer may potentially affect the accomplishment of the users’ learning objectives; and a learning design which merge declarative-level learning flow with instance-level data flow affects the reusability of the whole unit of learning (UoL). Based on the relation among these findings and literature, three dimensions of the IMS-LD data flow problem have been identified: the data flow automation problem already reported in the literature, which is related to the user’s data flow management issue; the data flow consistency problem, which is related to the issue of matching the different parts that comprise the data flow specification; and the UoL reuse problem, which is related to the instance-level collaborative data flow specification issue. Furthermore, these dimensions also help us to determine the necessary requirements in order to tackle these problems. Since IMS-LD fails to specify the data flow in collaborative learning, we propose a separation of the data flow from the learning flow. For this purpose, a standard workflow language such as BPEL is used and a unit of data flow (UoDF), which is just a business process archive (BPR) understandable by a BPEL-compliant engine is created. Then, the learning design is specified in a unit of learning flow (UoLF), which is actually a UoL understandable by an IMS-LD compliant engine that follows the best practices in collaborative data flow specification, which means not specifying the data flow at all. Finally, for coordination of both engines, a coordination model has been developed and a prototype is currently under test and evaluation. Future work includes evaluation of more case studies in order to validate the proposal solution, and identify limitations and drawbacks. Interestingly is that our proposal which is based on a composition-based approach may be thought as the in-between approach that will provide a framework for future integration of LD and workflow streams.