Keynote Speeches

Titles and abstracst are tentative, subject to change.

• Chair: Satoshi Fujita (Hiroshima University)
• Speaker: Koichi Wada (Hosei University)
• Title: Computational Power of Autonomous Mobile Robots with Lights—Boundary between Solvability and Unsolvability
• Abstract: In recent years, cooperation of a large number of autonomous mobile robots has received much attention. We consider algorithmic issues of autonomous mobile robots and its computational power on a theoretical model. In the model, each robot is modeled as a point in a plane and its capability is quite weak. It is usually assumed that robots are oblivious, anonymous and uniform. It is also assumed that each robot implicitly communicates with other robots by observing the environment including the positions of other robots. In the basic common model, most tasks cannot be solved without some additional assumptions. In this talk we consider visible external lights as an additional assumption and reveal the power of lights for autonomous mobile robots. In particular, we show some relationship between lights and other assumptions such as synchrony and restriction of robots moving and reveal boundary between solvability and unsovlability for gathering, in which all robots must meet in finite time at a point that is not predefined.

• Chair: Yasuhiko Nakashima (NAIST)
• Speaker: Kazuaki Ishizaki (IBM Research - Tokyo)
• Title : Transparent GPU exploitation for Java
• Abstract: For many applications such as analytics, machine learning, and deep learning, hardware accelerators that matter to achieve high performance. While several hardware accelerators are currently available, they require a programmer to explicitly write and optimize low-level operations. It is not easy for a non-expert programmer to use them. We are doing research to transparently exploit GPUs that are one of successful hardware accelerators, which avoids that a programmer explicitly writes low-level operations. In this talk, I will present two research projects: compiling a Java program for GPUs and exploiting GPUs in Apache Spark.

• Chair: Hiroshi Umeo (Osaka Electro-Communication University)
• Speaker: Rolf Hoffmann (Technische Universität Darmstadt)
• Title: Forming Target Patterns by Cellular Automata Agents
• Abstract: Considered is a 2D cellular automaton with moving agents. Each cell contains a particle with a certain spin/color that can be changed by an agent. The objective is to form a global target pattern, a special ordering of the colors, e.g. the forming of long strings of the same color. The quality of a target pattern is measured by the degree of order computed by counting the matching of 3 x 3 local patterns. Additional markers and signals between agents are used in order to improve the task's efficiency. The agents' behavior is controlled by a finite state machine (FSM). An agent can perform several actions as combinations of moving, turning, color changing, marker setting and signaling. It reacts on its own state and on the sensed colors, markers and signals. For a given training set of random initial configurations, near optimal FSMs were evolved by a genetic algorithm. The evolved agents are capable of forming target patterns with a certain degree of order. The performance of the system is evaluated depending on the number of agents and the using of the markers and signals.

• Chair: Kiyohito Yoshihara (KDDI Research, Inc.)
• Speaker: Makoto Suzuki (The University of Tokyo, Research Center for Advanced Science and Technology)
• Title: Wireless Sensor Networks for Smart Society
• Abstract: Wireless sensor networks play an important role for IoT system in various kinds of fields, such as agriculture, civil infrastructure, power plant, and so on. In this talk, we'll share our experience of the development and deployment of wireless sensor networks for tomato foliage monitoring, cable-stayed bridge monitoring, and urban expressway monitoring. Additionally, we present a fundamental sensor network technology motivated from the experience of these practical deployment.

• Chair: Tomoaki Tsumura (Nagoya Institute of Technology)
• Speaker: Akira Naruse (NVIDIA)
• Title: Accelerating training of large deep neural networks with the latest GPUs
• Abstract: Deep learning technologies have been advancing performance of many applications specifically related to image recognition or speech recognition in recent years. The keys to achieve good performance with deep learning are 1) advance of machine learning technologies, 2) big data and 3) enormous computing power. NVIDIA GPUs have high computing capability and have been supporting the emergence of deep learning as the back-end computing resources. These days, deep neural models are becoming larger and deeper in accordance with increase of data as we’re starting understand deeper models are necessary to achieve higher accuracy. Hence, more computing power is required. To respond to this request, some new features such as fp16 arithmetic, high bandwidth memory and NVLink are incorporated into the latest GPUs named Pascal architecture. In this session, we will give a detail explanation on how these features helping accelerating a training of larger and deeper neural network models.

• Chair: Jacir L. Bordim (University of Brasilia)
• Speaker : Ken-Ichi Ishikawa (Hiroshima University)
• Title : Accelerators in Lattice QCD simulations
• Abstract: Quantum Chromodynamics (QCD) is a fundamental theory describing the properties of protons and neutrons in terms of more fundamental particles (elementary particles), quarks and gluons. Protons and neutrons are composed of quarks and gluons and construct atomic nuclei with strongly interacting force. The properties of atomic nuclei must be described in terms of QCD. Because the interaction strength among quarks and quarks are too strong and complicated (non-linear) to apply the perturbation theory to QCD, it is difficult to analyze the properties analytically and theoretically. Lattice Quantum Chromodynamics (LQCD) is a nonperturbative method by which we can analyze QCD numerically. LQCD simulations has been done to investigate the properties of quarks and gluons using High Performance Computers(HPC) since 1980's and a couple of HPC machines has been developed for LQCD. The most time consuming part of LQCD simulations is to solve the equation of motion of quarks, the Dirac equation, discretized on a four-dimensional space-time lattice box. The equation becomes a large scale linear equation with a sparse matrix. To solve it the iterative solver algorithms are employed typically. The LQCD now succeeds in describing the properties of single proton or nutron, and is progressing to investigate the properties of the muti-proton/nutron systems (=atomic nuclei). In recent years the application of accelerators to LQCD simulations becomes common and a key technology. The pioneering GPU application (GPGPU) to LQCD has been done using OpenGL by Egri et al. in 2007 before CUDA, and CUDA is now common language to utilize GPU accelerator to LQCD. Intel Xeon Phi is a new accelerator architecture and is also optimized to LQCD. The equation of motion of quarks is solved by using these accelerators. In this talk I review the LQCD method and the current status of the utilization of accelerators to LQCD simulations, where CUDA GPU and Xeon Phi will be presented.

• Speaker: Vladimir Voevodin (Moscow State University)
• Title: Algorithmic Challenges of Legacy HPC Application Migration
• Abstract: High performance computing clusters, SMP-machines, GPU- or Phi-based computing systems, vector computers, FPGA accelerators, small, medium or petascale supercomputers. Diversity of modern computing systems is great and their huge potential allows complex problems, previously thought impossible, to be solved. By now, computing community has developed highly valuable and wide spectrum of software with a hope to use this wealth of codes on the next generation computers for a long time. At the same time, the efficient usage of all opportunities offered by modern computing systems through a large number of applications represents a global challenge and requires new knowledge, skills and abilities, where one of the main roles belongs to understanding of key properties of parallel algorithms. The talk will address the urgent need for theoretical and practical technologies of an accurate and concerted design of highly parallel algorithms and extreme scaled applications to be able to solve large problems through a variety of different computing platforms. Deep understanding of algorithmic structures will provide a key to successful application migration while preserving a high level of efficiency.

• Chair: Susumu Matsumae (Saga University)
• Speaker: Yukiko Yamauchi (Kyushu University)
• Title: Self-organization of Mobile Robots and Rotation Groups
• Abstract: Distributed coordination of autonomous mobile robots attracts much attention because of its wide application for wheeled robots, drones, molecular robots, and so on. We consider the theoretical aspect of self-organization of mobile robots moving in the three-dimensional space (3D-space). The robots are anonymous, oblivious (memory-less), uniform (i.e., executes the same algorithm), and have neither any access to the global coordinate system nor any explicit communication medium. The formation problem requires the robots to form a given shape such as a point, a circle, a plane, and an arbitrary target pattern. Because the robots cannot break the rotational symmetry among themselves, the formation power of a robot system depends on the symmetry of the initial configuration of the robots. In 3D-space, rotational symmetry is classified into five types of rotation groups, and the formable patterns are characterized by using them. In this talk, we survey recent results on the formation problems in 3D-space, that have been built on top of the results on the formation problems in 2D-space.

• Chair: Yasuyuki Nogami (Okayama University)
• Speaker: Toshihiro Yamauchi (Okayama University)
• Title: Recent Topics on Use-after-free Exploitation and Mitigation Techniques
• Abstract: Use-After-Free (UAF) vulnerabilities, which are abused by exploiting a dangling pointer that refers to a freed memory, have increased. In particular, large-scale programs such as browsers often include many dangling pointers, and UAF vulnerabilities are frequently exploited by drive-by download attacks. In this presentation, I will talk about therecent UAF attack and the mitigation techniques, and presents our UAF-attack prevention method that prevents the reuse of freed memory area.

• Chair: Shuichi Ichikawa (Toyohashi University of Technology)
• Speaker: Jacir L. Bordim (University of Brasilia)
• Title: Dynamic Spectrum Access Networks
• Abstract: Owing to its flexibility, low-cost and easy-deployment, wireless technologies have become commonplace in a wide-range of consumer electronics, such as smartphones, cameras and e-readers. On the other hand, Medium Access Control (MAC) protocols strive to deal with an increasing number of contending nodes on crowded ISM (Industrial, Scientific and Medical) frequency bands. Conversely, licensed bands are reported to be mostly underutilized. In this context, dynamic spectrum access appears as an alternative to improve spectrum allocation by relaxing the rules on licensed spectrum. However, one of the main problems faced in spectrum reuse is the difficulty unlicensed users have to explore spectrum opportunities. This keynote talk addresses some of the issues faced on dynamic spectrum allocation and points out some of the research challenges and opportunities in the field.

• Speaker: Hiroshi Inoue (IBM Research – Tokyo)
• Title: Efficient Optimization of Diameter and Average Shortest Path Length of a Graph using Path Count Index
• Abstract: In this talk, we present an efficient method to find an undirected and unweighted graph having a smaller diameter and a shorter average shortest path length (ASPL) with given order and degree. For this problem, local search is one of the most common techniques to find a better solution. However, due to the high computation cost of the objective functions (diameter and ASPL of a graph), the naive local search is not practical for large graphs. Our proposed technique reduces the computation cost drastically by using alternative objective functions that estimate the real objective functions. The key for the higher performance is that what we actually need for the local search is not the diameter and ASPL of the graph, but the changes of them due to a small modification applied for the graph; the changes can be calculated using only the local information around the modified edges without involving the whole graph. To calculate the changes without fully recomputing the node-to-node distances, a new data structure called Path Count Index is introduced. Our technique accelerated the local search significantly (by order of 105 in maximum). Although the Path Count Index consumed additional memory space, it was possible to execute the local search for graphs with up to 10,000 nodes on a commodity PC with 4 GB of system memory.