Matteo Turilli

Shantenu Jha, Vincent R. Pascuzzi, Matteo Turilli

Increasingly, scientific discovery requires sophisticated and scalable workflows. Workflows have become the ``new applications,'' wherein multi-scale computing campaigns comprise multiple and heterogeneous executable tasks. In particular, the introduction of AI/ML models into the traditional HPC workflows has been an enabler of highly accurate modeling, typically reducing computational needs compared to traditional methods. This chapter discusses various modes of integrating AI/ML models to HPC computations, resulting in diverse types of AI-coupled HPC workflows. The increasing need of coupling AI/ML and HPC across scientific domains is motivated, and then exemplified by a number of production-grade use cases for each mode. We additionally discuss the primary challenges of extreme-scale AI-coupled HPC campaigns -- task heterogeneity, adaptivity, performance -- and several framework and middleware solutions which aim to address them. While both HPC workflow and AI/ML computing paradigms are independently effective, we highlight how their integration, and ultimate convergence, is leading to significant improvements in scientific performance across a range of domains, ultimately resulting in scientific explorations otherwise unattainable.

Vincent R. Pascuzzi, Ozgur O. Kilic, Matteo Turilli et al.

Heterogeneous scientific workflows consist of numerous types of tasks that require executing on heterogeneous resources. Asynchronous execution of those tasks is crucial to improve resource utilization, task throughput and reduce workflows' makespan. Therefore, middleware capable of scheduling and executing different task types across heterogeneous resources must enable asynchronous execution of tasks. In this paper, we investigate the requirements and properties of the asynchronous task execution of machine learning (ML)-driven high performance computing (HPC) workflows. We model the degree of asynchronicity permitted for arbitrary workflows and propose key metrics that can be used to determine qualitative benefits when employing asynchronous execution. Our experiments represent relevant scientific drivers, we perform them at scale on Summit, and we show that the performance enhancements due to asynchronous execution are consistent with our model.

Aymen Alsaadi, Jonathan Ash, Mikhail Titov et al.

Computational protein design is experiencing a transformation driven by AI/ML. However, the range of potential protein sequences and structures is astronomically vast, even for moderately sized proteins. Hence, achieving convergence between generated and predicted structures demands substantial computational resources for sampling. The Integrated Machine-learning for Protein Structures at Scale (IMPRESS) offers methods and advanced computing systems for coupling AI to high-performance computing tasks, enabling the ability to evaluate the effectiveness of protein designs as they are developed, as well as the models and simulations used to generate data and train models. This paper introduces IMPRESS and demonstrates the development and implementation of an adaptive protein design protocol and its supporting computing infrastructure. This leads to increased consistency in the quality of protein design and enhanced throughput of protein design due to dynamic resource allocation and asynchronous workload execution.

Andre Merzky, Mikhail Titov, Matteo Turilli et al.

Hybrid workflows combining traditional HPC and novel ML methodologies are transforming scientific computing. This paper presents the architecture and implementation of a scalable runtime system that extends RADICAL-Pilot with service-based execution to support AI-out-HPC workflows. Our runtime system enables distributed ML capabilities, efficient resource management, and seamless HPC/ML coupling across local and remote platforms. Preliminary experimental results show that our approach manages concurrent execution of ML models across local and remote HPC/cloud resources with minimal architectural overheads. This lays the foundation for prototyping three representative data-driven workflow applications and executing them at scale on leadership-class HPC platforms.

Austin Clyde, Thomas Brettin, Alexander Partin et al.

We propose a benchmark to study surrogate model accuracy for protein-ligand docking. We share a dataset consisting of 200 million 3D complex structures and 2D structure scores across a consistent set of 13 million "in-stock" molecules over 15 receptors, or binding sites, across the SARS-CoV-2 proteome. Our work shows surrogate docking models have six orders of magnitude more throughput than standard docking protocols on the same supercomputer node types. We demonstrate the power of high-speed surrogate models by running each target against 1 billion molecules in under a day (50k predictions per GPU seconds). We showcase a workflow for docking utilizing surrogate ML models as a pre-filter. Our workflow is ten times faster at screening a library of compounds than the standard technique, with an error rate less than 0.01\% of detecting the underlying best scoring 0.1\% of compounds. Our analysis of the speedup explains that to screen more molecules under a docking paradigm, another order of magnitude speedup must come from model accuracy rather than computing speed (which, if increased, will not anymore alter our throughput to screen molecules). We believe this is strong evidence for the community to begin focusing on improving the accuracy of surrogate models to improve the ability to screen massive compound libraries 100x or even 1000x faster than current techniques.

Alexander Brace, Igor Yakushin, Heng Ma et al.

Machine learning (ML)-based steering can improve the performance of ensemble-based simulations by allowing for online selection of more scientifically meaningful computations. We present DeepDriveMD, a framework for ML-driven steering of scientific simulations that we have used to achieve orders-of-magnitude improvements in molecular dynamics (MD) performance via effective coupling of ML and HPC on large parallel computers. We discuss the design of DeepDriveMD and characterize its performance. We demonstrate that DeepDriveMD can achieve between 100-1000x acceleration for protein folding simulations relative to other methods, as measured by the amount of simulated time performed, while covering the same conformational landscape as quantified by the states sampled during a simulation. Experiments are performed on leadership-class platforms on up to 1020 nodes. The results establish DeepDriveMD as a high-performance framework for ML-driven HPC simulation scenarios, that supports diverse MD simulation and ML back-ends, and which enables new scientific insights by improving the length and time scales accessible with current computing capacity.

Agastya P. Bhati, Shunzhou Wan, Dario Alfè et al.

The race to meet the challenges of the global pandemic has served as a reminder that the existing drug discovery process is expensive, inefficient and slow. There is a major bottleneck screening the vast number of potential small molecules to shortlist lead compounds for antiviral drug development. New opportunities to accelerate drug discovery lie at the interface between machine learning methods, in this case developed for linear accelerators, and physics-based methods. The two in silico methods, each have their own advantages and limitations which, interestingly, complement each other. Here, we present an innovative infrastructural development that combines both approaches to accelerate drug discovery. The scale of the potential resulting workflow is such that it is dependent on supercomputing to achieve extremely high throughput. We have demonstrated the viability of this workflow for the study of inhibitors for four COVID-19 target proteins and our ability to perform the required large-scale calculations to identify lead antiviral compounds through repurposing on a variety of supercomputers.

Hyungro Lee, Heng Ma, Matteo Turilli et al.

Simulations of biological macromolecules play an important role in understanding the physical basis of a number of complex processes such as protein folding. Even with increasing computational power and evolution of specialized architectures, the ability to simulate protein folding at atomistic scales still remains challenging. This stems from the dual aspects of high dimensionality of protein conformational landscapes, and the inability of atomistic molecular dynamics (MD) simulations to sufficiently sample these landscapes to observe folding events. Machine learning/deep learning (ML/DL) techniques, when combined with atomistic MD simulations offer the opportunity to potentially overcome these limitations by: (1) effectively reducing the dimensionality of MD simulations to automatically build latent representations that correspond to biophysically relevant reaction coordinates (RCs), and (2) driving MD simulations to automatically sample potentially novel conformational states based on these RCs. We examine how coupling DL approaches with MD simulations can fold small proteins effectively on supercomputers. In particular, we study the computational costs and effectiveness of scaling DL-coupled MD workflows by folding two prototypical systems, viz., Fs-peptide and the fast-folding variant of the villin head piece protein. We demonstrate that a DL driven MD workflow is able to effectively learn latent representations and drive adaptive simulations. Compared to traditional MD-based approaches, our approach achieves an effective performance gain in sampling the folded states by at least 2.3x. Our study provides a quantitative basis to understand how DL driven MD simulations, can lead to effective performance gains and reduced times to solution on supercomputing resources.

Vivek Balasubramanian, Shantenu Jha, Andre Merzky et al.

RADICAL-Cybertools (RCT) are a set of software systems that serve as middleware to develop efficient and effective tools for scientific computing. Specifically, RCT enable executing many-task applications at extreme scale and on a variety of computing infrastructures. RCT are building blocks, designed to work as stand-alone systems, integrated among themselves or integrated with third-party systems. RCT enables innovative science in multiple domains, including but not limited to biophysics, climate science and particle physics, consuming hundreds of millions of core hours. This paper provides an overview of RCT systems, their impact, and the architectural principles and software engineering underlying RCT

Danila Oleynik, Sergey Panitkin, Matteo Turilli et al.

The computing systems used by LHC experiments has historically consisted of the federation of hundreds to thousands of distributed resources, ranging from small to mid-size resource. In spite of the impressive scale of the existing distributed computing solutions, the federation of small to mid-size resources will be insufficient to meet projected future demands. This paper is a case study of how the ATLAS experiment has embraced Titan---a DOE leadership facility in conjunction with traditional distributed high- throughput computing to reach sustained production scales of approximately 52M core-hours a years. The three main contributions of this paper are: (i) a critical evaluation of design and operational considerations to support the sustained, scalable and production usage of Titan; (ii) a preliminary characterization of a next generation executor for PanDA to support new workloads and advanced execution modes; and (iii) early lessons for how current and future experimental and observational systems can be integrated with production supercomputers and other platforms in a general and extensible manner.

Matteo Turilli, Andre Merzky, Vivek Balasubramanian et al.

We suggest there is a need for a fresh perspective on the design and development of workflow systems and argue for a building blocks approach. We outline a description of this approach and define the properties of software building blocks. We discuss RADICAL-Cybertools as one implementation of the building blocks concept, showing how they have been designed and developed in accordance with this approach. Four case studies are presented, covering a dozen science problems. We discuss how RADICAL-Cybertools have been used to develop new workflow systems capabilities and integrated to enhance existing ones, illustrating the applicability and potential of software building blocks. In doing so, we have begun an investigation of an alternative approach to thinking about the design and implementation of workflow systems.

Matteo Turilli, Mark Santcroos, Shantenu Jha

Pilot-Job systems play an important role in supporting distributed scientific computing. They are used to consume more than 700 million CPU hours a year by the Open Science Grid communities, and by processing up to 1 million jobs a day for the ATLAS experiment on the Worldwide LHC Computing Grid. With the increasing importance of task-level parallelism in high-performance computing, Pilot-Job systems are also witnessing an adoption beyond traditional domains. Notwithstanding the growing impact on scientific research, there is no agreement upon a definition of Pilot-Job system and no clear understanding of the underlying abstraction and paradigm. Pilot-Job implementations have proliferated with no shared best practices or open interfaces and little interoperability. Ultimately, this is hindering the realization of the full impact of Pilot-Jobs by limiting their robustness, portability, and maintainability. This paper offers a comprehensive analysis of Pilot-Job systems critically assessing their motivations, evolution, properties, and implementation. The three main contributions of this paper are: (i) an analysis of the motivations and evolution of Pilot-Job systems; (ii) an outline of the Pilot abstraction, its distinguishing logical components and functionalities, its terminology, and its architecture pattern; and (iii) the description of core and auxiliary properties of Pilot-Jobs systems and the analysis of seven exemplar Pilot-Job implementations. Together, these contributions illustrate the Pilot paradigm, its generality, and how it helps to address some challenges in distributed scientific computing.