arXiv subjects

Maksim Levental

Publications and source records attributed to Maksim Levental.

11 recordsLinked to original sources

arXiv2024

Trillion Parameter AI Serving Infrastructure for Scientific Discovery: A Survey and Vision

Deep learning methods are transforming research, enabling new techniques, and ultimately leading to new discoveries. As the demand for more capable AI models continues to grow, we are now entering an era of Trillion Parameter Models (TPM), or models with more than a trillion parameters -- such as Huawei's PanGu-$Σ$. We describe a vision for the ecosystem of TPM users and providers that caters to the specific needs of the scientific community. We then outline the significant technical challenges and open problems in system design for serving TPMs to enable scientific research and discovery. Specifically, we describe the requirements of a comprehensive software stack and interfaces to support the diverse and flexible requirements of researchers.

cs.LG↗

arXiv2023

nelli: a lightweight frontend for MLIR

Multi-Level Intermediate Representation (MLIR) is a novel compiler infrastructure that aims to provide modular and extensible components to facilitate building domain specific compilers. However, since MLIR models programs at an intermediate level of abstraction, and most extant frontends are at a very high level of abstraction, the semantics and mechanics of the fundamental transformations available in MLIR are difficult to investigate and employ in and of themselves. To address these challenges, we have developed \texttt{nelli}, a lightweight, Python-embedded, domain-specific, language for generating MLIR code. \texttt{nelli} leverages existing MLIR infrastructure to develop Pythonic syntax and semantics for various MLIR features. We describe \texttt{nelli}'s design goals, discuss key details of our implementation, and demonstrate how \texttt{nelli} enables easily defining and lowering compute kernels to diverse hardware platforms.

cs.PL↗

arXiv2023

OpenHLS: High-Level Synthesis for Low-Latency Deep Neural Networks for Experimental Science

In many experiment-driven scientific domains, such as high-energy physics, material science, and cosmology, high data rate experiments impose hard constraints on data acquisition systems: collected data must either be indiscriminately stored for post-processing and analysis, thereby necessitating large storage capacity, or accurately filtered in real-time, thereby necessitating low-latency processing. Deep neural networks, effective in other filtering tasks, have not been widely employed in such data acquisition systems, due to design and deployment difficulties. We present an open source, lightweight, compiler framework, without any proprietary dependencies, OpenHLS, based on high-level synthesis techniques, for translating high-level representations of deep neural networks to low-level representations, suitable for deployment to near-sensor devices such as field-programmable gate arrays. We evaluate OpenHLS on various workloads and present a case-study implementation of a deep neural network for Bragg peak detection in the context of high-energy diffraction microscopy. We show OpenHLS is able to produce an implementation of the network with a throughput 4.8 $μ$s/sample, which is approximately a 4$\times$ improvement over the existing implementation

cs.AR↗

arXiv2022

Real-Time Streaming and Event-driven Control of Scientific Experiments

Advancements in scientific instrument sensors and connected devices provide unprecedented insight into ongoing experiments and present new opportunities for control, optimization, and steering. However, the diversity of sensors and heterogeneity of their data result in make it challenging to fully realize these new opportunities. Organizing and synthesizing diverse data streams in near-real-time requires both rich automation and Machine Learning (ML). To efficiently utilize ML during an experiment, the entire ML lifecycle must be addressed, including refining experiment configurations, retraining models, and applying decisions-tasks that require an equally diverse array of computational resources spanning centralized HPC to the accelerators at the edge. Here we present the Manufacturing Data and Machine Learning platform (MDML). The MDML is designed to standardize the research and operational environment for advanced data analytics and ML-enabled automated process optimization by providing the cyberinfrastructure to integrate sensor data streams and AI in cyber-physical systems for in-situ analysis. To achieve this, the MDML provides a fabric to receive and aggregate IoT data and simultaneously orchestrate remote computation across the computing continuum. In this paper we describe the MDML and show how it is used in advanced manufacturing to act on IoT data and orchestrate distributed ML to guide experiments.

cs.DC↗

arXiv2022

Ultrafast Focus Detection for Automated Microscopy

Technological advancements in modern scientific instruments, such as scanning electron microscopes (SEMs), have significantly increased data acquisition rates and image resolutions enabling new questions to be explored; however, the resulting data volumes and velocities, combined with automated experiments, are quickly overwhelming scientists as there remain crucial steps that require human intervention, for example reviewing image focus. We present a fast out-of-focus detection algorithm for electron microscopy images collected serially and demonstrate that it can be used to provide near-real-time quality control for neuroscience workflows. Our technique, \textit{Multi-scale Histologic Feature Detection}, adapts classical computer vision techniques and is based on detecting various fine-grained histologic features. We exploit the inherent parallelism in the technique to employ GPU primitives in order to accelerate characterization. We show that our method can detect of out-of-focus conditions within just 20ms. To make these capabilities generally available, we deploy our feature detector as an on-demand service and show that it can be used to determine the degree of focus in approximately 230ms, enabling near-real-time use.

eess.IV↗

arXiv2022

Memory Planning for Deep Neural Networks

We study memory allocation patterns in DNNs during inference, in the context of large-scale systems. We observe that such memory allocation patterns, in the context of multi-threading, are subject to high latencies, due to \texttt{mutex} contention in the system memory allocator. Latencies incurred due to such \texttt{mutex} contention produce undesirable bottlenecks in user-facing services. Thus, we propose a "memorization" based technique, \texttt{MemoMalloc}, for optimizing overall latency, with only moderate increases in peak memory usage. Specifically, our technique consists of a runtime component, which captures all allocations and uniquely associates them with their high-level source operation, and a static analysis component, which constructs an efficient allocation "plan". We present an implementation of \texttt{MemoMalloc} in the PyTorch deep learning framework and evaluate memory consumption and execution performance on a wide range of DNN architectures. We find that \texttt{MemoMalloc} outperforms state-of-the-art general purpose memory allocators, with respect to DNN inference latency, by as much as 40\%.

cs.LG↗

arXiv2021

A Hardware Co-design Workflow for Scientific Instruments at the Edge

As spatial and temporal resolutions of scientific instruments improve, the explosion in the volume of data produced is becoming a key challenge. It can be a critical bottleneck for integration between scientific instruments at the edge and high-performance computers/emerging accelerators. Placing data compression or reduction logic close to the data source is a possible approach to solve the bottleneck. However, the realization of such a solution requires the development of custom ASIC designs, which is still challenging in practice and tends to produce one-off implementations unusable beyond the initial intended scope. Therefore, as a feasibility study, we have been investigating a design workflow that allows us to explore algorithmically complex hardware designs and develop reusable hardware libraries for the needs of scientific instruments at the edge. Our vision is to cultivate our hardware development capability for streaming/dataflow hardware components that can be placed close to the data source to enable extreme data-intensive scientific experiments or environmental sensing. Furthermore, reducing data movement is essential to improving computing performance in general. Therefore, our co-design efforts on streaming hardware components can benefit computing applications other than scientific instruments. This vision paper discusses hardware specialization needs in scientific instruments and briefly reviews our progress leveraging the Chisel hardware description language and emerging open-source hardware ecosystems, including a few design examples.

physics.ins-det↗

arXiv2021

Tensor Networks for Simulating Quantum Circuits on FPGAs

Most research in quantum computing today is performed against simulations of quantum computers rather than true quantum computers. Simulating a quantum computer entails implementing all of the unitary operators corresponding to the quantum gates as tensors. For high numbers of qubits, performing tensor multiplications for these simulations becomes quite expensive, since $N$-qubit gates correspond to $2^{N}$-dimensional tensors. One way to accelerate such a simulation is to use field programmable gate array (FPGA) hardware to efficiently compute the matrix multiplications. Though FPGAs can efficiently perform tensor multiplications, they are memory bound, having relatively small block random access memory. One way to potentially reduce the memory footprint of a quantum computing system is to represent it as a tensor network; tensor networks are a formalism for representing compositions of tensors wherein economical tensor contractions are readily identified. Thus we explore tensor networks as a means to reducing the memory footprint of quantum computing systems and broadly accelerating simulations of such systems.

quant-ph↗

arXiv2021

Accelerated, Scalable and Reproducible AI-driven Gravitational Wave Detection

The development of reusable artificial intelligence (AI) models for wider use and rigorous validation by the community promises to unlock new opportunities in multi-messenger astrophysics. Here we develop a workflow that connects the Data and Learning Hub for Science, a repository for publishing AI models, with the Hardware Accelerated Learning (HAL) cluster, using funcX as a universal distributed computing service. Using this workflow, an ensemble of four openly available AI models can be run on HAL to process an entire month's worth (August 2017) of advanced Laser Interferometer Gravitational-Wave Observatory data in just seven minutes, identifying all four all four binary black hole mergers previously identified in this dataset and reporting no misclassifications. This approach combines advances in AI, distributed computing, and scientific data infrastructure to open new pathways to conduct reproducible, accelerated, data-driven discovery.

gr-qc↗

arXiv2020

Comparing the costs of abstraction for DL frameworks

High level abstractions for implementing, training, and testing Deep Learning (DL) models abound. Such frameworks function primarily by abstracting away the implementation details of arbitrary neural architectures, thereby enabling researchers and engineers to focus on design. In principle, such frameworks could be "zero-cost abstractions"; in practice, they incur translation and indirection overheads. We study at which points exactly in the engineering life-cycle of a DL model the highest costs are paid and whether they can be mitigated. We train, test, and evaluate a representative DL model using PyTorch, LibTorch, TorchScript, and cuDNN on representative datasets, comparing accuracy, execution time and memory efficiency.

cs.LG↗

arXiv2020

Towards Online Steering of Flame Spray Pyrolysis Nanoparticle Synthesis

Flame Spray Pyrolysis (FSP) is a manufacturing technique to mass produce engineered nanoparticles for applications in catalysis, energy materials, composites, and more. FSP instruments are highly dependent on a number of adjustable parameters, including fuel injection rate, fuel-oxygen mixtures, and temperature, which can greatly affect the quality, quantity, and properties of the yielded nanoparticles. Optimizing FSP synthesis requires monitoring, analyzing, characterizing, and modifying experimental conditions.Here, we propose a hybrid CPU-GPU Difference of Gaussians (DoG)method for characterizing the volume distribution of unburnt solution, so as to enable near-real-time optimization and steering of FSP experiments. Comparisons against standard implementations show our method to be an order of magnitude more efficient. This surrogate signal can be deployed as a component of an online end-to-end pipeline that maximizes the synthesis yield.

cs.CV↗