Heterogeneity in Hardware: Opportunities and Challenges for Software and Applications

Abstract

With the hardware technology scaling and the trend on heterogeneous chip design, the exis

We present Exascale GeoStatistics (ExaGeoStat) software, a high-performance library implemented on a wide variety of contemporary hybrid distributed-shared supercomputers whose primary target is climate and environmental prediction applications.

Out-of-Core simulation systems often produce a massive amount of data that cannot fit on the aggregate fast memory of the compute nodes, and they also require to read back these data for computation. As a result, I/O data movement can be a bottleneck in large-scale simulations. Advances in memory architecture have made it feasible to integrate hierarchical storage media on large-scale systems, starting from the traditional Parallel File Systems to intermediate fast disk technologies (e.g., node-local and remote-shared NVMe and SSD-based Burst Buffers) and up to CPU’s main memory and GPU’s High Bandwidth Memory. However, while adding additional and faster storage media increases I/O bandwidth, it pressures the CPU, as it becomes responsible for managing and moving data between these layers of storage. Simulation systems are thus vulnerable to being blocked by I/O operations. The Multilayer Buffer System (MLBS) proposed in this research demonstrates a general method for overlapping I/O with computation that helps to ameliorate the strain on the processors through asynchronous access. The main idea consists in decoupling I/O operations from computational phases using dedicated hardware resources to perform expensive context switches. By continually prefetching up and down across all hardware layers of the memory/storage subsystems, MLBS transforms the original I/O-bound behavior of evaluated applications and shifts it closer to a memory-bound or compute-bound regime.

In the lecture we present a three dimensional mdoel for the simulation of signal processing in neurons. To handle problems of this complexity, new mathematical methods and software tools are required. In recent years, new approaches such as parallel adaptive multigrid methods and corresponding software tools have been developed allowing to treat problems of huge complexity. Part of this approach is a method to reconstruct the geometric structure of neurons from data measured by 2-photon microscopy. Being able to reconstruct neural geometries and network connectivities from measured data is the basis of understanding coding of motoric perceptions and long term plasticity which is one of the main topics of neuroscience. Other issues are compartment models and upscaling.

In this talk, we review some recent advances in the analysis and design of algebraic flux correction (AFC) schemes for hyperbolic problems. In contrast to most variational stabilization techniques, AFC approaches modify the standard Galerkin discretization in a way which provably guarantees the validity of discrete maximum principles for scalar conservation laws and invariant domain preservation for hyperbolic systems. The corresponding inequality constraints are enforced by adding diffusive fluxes, and bound-preserving antidiffusive corrections are performed to obtain nonlinear high-order approximations. After introducing the AFC methodology and the underlying theoretical framework in the context of continuous piecewise-linear finite element discretizations, we present some of the limiting techniques that we use in high-resolution AFC schemes. This presentation is based on joint work with Dr. Manuel Quezada de Luna (KAUST) and other collaborators.

The SLATE (Software for Linear Algebra Targeting Exascale) library is being developed to provide fundamental dense linear algebra capabilities for current and upcoming distributed high-performance systems, both accelerated CPU–GPU based and CPU based.

This talk will be a gentle introduction to proximal splitting algorithms to minimize a sum of possibly nonsmooth convex functions. Several such algorithms date back to the 60s, but the last 10 years have seen the development of new primal-dual splitting algorithms, motivated by the need to solve large-scale problems in signal and image processing, machine learning, and more generally data science. No background will be necessary to attend the talk, whose goal is to present the intuitions behind this class of methods.

In this talk, I will describe two recent results on detecting and understanding backdoor attacks on deep learning systems. I will first present Neural Cleanse (IEEE S&P 2019), the first robust tool to detect a wide range of backdoors in deep learning models. We use the idea of perturbation distances between classification labels to detect when a backdoor trigger has created shortcuts to misclassification to a particular label.  Second, I will also summarize our new work on Latent Backdoors (CCS 2019), a stronger type of backdoor attack that is more difficult to detect and survives retraining in commonly used transfer learning systems. Latent backdoors are robust and stealthy, even against the latest detection tools (including neural cleanse).

We consider the problem of recovering a low-rank signal matrix in the presence of a general, unknown additive noise; more specifically, noise where the eigenvalues of the sample covariance matrix have a general bulk distribution. We assume given an upper bound for the rank of the assumed orthogonally invariant signal, and develop a selector for hard thresholding of singular values, which adapts to the unknown correlation structure of the noise.

A variety of intriguing patterns in eigenvalues were observed and speculated about in ML conference papers. We describe the work of Vardan Papyan showing that the traditional subdisciplines, properly deployed, can offer insights about these objects that ML researchers had.

Adil Salim is mainly interested in stochastic approximation, optimization, and machine learning. He is currently a Postdoctoral Research Fellow working with Professor Peter Richtarik at the Visual Computing Center (VCC) at King Abdullah University of Science and Technology (KAUST).

Abstract

Many physical models are characteri

In this talk, Sumayah will survey a number of malicious hosting infrastructures for different services and approaches to detecting them. Among them are works on an emerging trend of Bulletproof hosting services reselling infrastructure from lower-end service providers, use of residential proxy as a service to avoid server-side blocking and DNS based hosting infrastructure.

This talk will try to clarify some misconceptions about what digital health (DH) is, and what it should not be.

Computing and communications infrastructures have become commodities that transact huge quantities of data and are pervasively interconnected, inside countries, and worldwide. Modern societies largely depend on them.

Existing RDF engines are designed for specific hardware architectures; porting to a different architecture (e.g., GPUs) entails enormous implementation effort. We explore sparse matrix algebra as an alternative for designing a portable, scalable and efficient RDF engine.

Abstract

The talk will present the integrate

Current computing infrastructures are evolving from traditional systems to environments that involve sensors, edge devices, instruments, and high-end  computing power in the cloud and HPC systems. A key aspect is how to describe the applications to be executed in such platforms. Very often these applications are not standalone, but involve a set of sub-applications or steps composing a workflow. The scientists then rely on effective environments to describe their workflows and engines to manage them in complex infrastructures. Refreshments - Brown Bag Lunch will be served.

Materials science is a multidisciplinary field of research with many different scientists and engineers having various backgrounds active in it. The literature landscape consequently is populated currently by a wide range of journals which greatly differ in purpose, scope, quality, and readership. Jos Lenders, Deputy Editor of Advanced Materials, Advanced Functional Materials, and Advanced Optical Materials, will track some of the most important developments and trends in the research field and the Advanced journals program. Last year, Advanced Materials reached an Impact Factor of 21.95 and received over 8,300 submissions – and Advanced Functional Materials over 9,200. Only around 15% of all those papers made it to publication in the journal, and this rate is similar for all other Advanced journals. So, what do editors do to select the very best papers, and what can authors do to optimize their chances of having their manuscripts accepted?

This series of lectures guide students to the preparation and analysis of a well-organized abstract. We will discuss the proper language (tense, voice, and person) for abstract writing, and learn how to meet the purposes of different abstracts. Finally, students will have a chance to compose and evaluate their writing. Topics: Overview of abstract writing; Conference abstract journal abstract; Organization of an abstract; Language conventions of abstract writing; Disciplinary abstract analysis; Frequent mistakes of abstract writing.

Many problems in machine learning rely on statistics and optimization. To solve these problems, new techniques are needed. I will show some of these new techniques through selected machine learning problems I have recently worked on, such as nonconvex stochastic optimization, distributed training, adversarial attack, and generative models.

Designing a secure permissionless distributed ledger that performs on par with centralized payment processors such as Visa is challenging. Most existing distributed ledgers are unable to "scale-out" -- growing total processing capacity with number of participants -- and those that do compromise security or decentralization. This work presents OmniLedger, the first scale-out distributed ledger that can preserve long-term security under permissionless operation. OmniLedger ensures strong correctness and security by using a bias-resistant public randomness protocol to choose large statistically representative shards to process transactions, and by introducing an efficient cross-shard commit protocol to handle transactions affecting multiple shards atomically.

This talk begins with an introduction to quantum chemistry for HPC researchers.  We discuss the main computational kernels, how their performance is limited by various bottlenecks, and algorithms and implementations that improve performance.  One particular kernel is the calculation of the Coulomb matrix.  Whereas the fast multipole method (FMM) can be used to rapidly compute the Coulomb potential for sets of point charges, continuous variants of FMM were developed in the 1990s for sets of charge distributions that arise in quantum chemistry.