Performance characterization of global address space applications: a case study with nwchem
Abstract not provided
SEARCH · Search NASA
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
Abstract not provided
As parallel applications running on large–scale computing systems become increasingly memory constrained, the ability to attribute memory usage to the various components of the application is becoming increasingly important. We present the design and implementation of memnesia, a novel memory usage profiler for parallel and distributed message–passing applications. Our approach captures both application– and message–passing library–specific memory usage statistics from unmodified binaries dynamically linked to a message–passing communication library. Using microbenchmarks and proxy applications, we evaluated our profiler across three Message Passing Interface (MPI) implementations and two hardware platforms. Furthermore, the results show that our approach and the corresponding implementation can accurately quantify memory resource usage as a function of time, scale, communication workload, and software or hardware system architecture, clearly distinguishing between application and MPI library memory usage at a per–process level. With this new capability, we show that job size, communication workload, and hardware/software architecture influence peak runtime memory usage. In practice, this tool provides a potentially valuable source of information for application developers seeking to measure and optimize memory usage.
An emerging trend in High Performance Computing (HPC) systems that use hierarchical topologies (such as dragonfly) is that the applications are increasingly exhibiting high run-to-run performance variability. This poses a significant challenge for application developers, job schedulers, and system maintainers. One approach to address the performance variability is to use newly proposed network topologies such as megafly (or dragonfly+) that offer increased path diversity compared to a traditional fully connected dragonfly. Yet another approach is to use quality of service (QoS) traffic classes that ensure bandwidth guarantees. In this work, we select HPC application workloads that have exhibited performance variability on current 2-D dragonfly systems. We evaluate the baseline performance expectations of these workloads on megafly and 1-D dragonfly network models with comparably similar network configurations. Our results show that the megafly network, despite using fewer virtual channels (VCs) for deadlock avoidance than a dragonfly, performs as well as a fully connected 1-D dragonfly network. We then exploit the fact that megafly networks require fewer VCs to incorporate QoS traffic classes. We use bandwidth capping and traffic differentiation techniques to introduce multiple traffic classes in megafly networks. In some cases, our results show that QoS can completely mitigate application performance variability while causing minimal slowdown to the background network traffic.
New experimental and AI-driven workloads are moving into the realm of extreme-scale HPC systems at the same time that high-performance flash is becoming cost-effective to deploy at scale. This confluence poses a number of new technical and economic challenges and opportunities in designing the next generation of HPC storage and I/O subsystems to achieve the right balance of bandwidth, latency, endurance, and cost. In this work, we present quantitative models that use workload data from existing, disk-based file systems to project the architectural requirements of all-flash Lustre file systems. Using data from NERSC’s Cori I/O subsystem, we then demonstrate the minimum required capacity for data, capacity for metadata and data-on-MDT, and SSD endurance for a future all-flash Lustre file system.
The Energy Exascale Earth System Model (E3SM) Land Model (ELM) has been extended to kilometer-scale (km-ELM) resolutions, enabling high-fidelity simulations of terrestrial processes at 1 km x 1 km grid spacing. In ELM, domain decomposition partitions the computational domain across processors, ensuring efficient parallel execution. Currently, round-robin decomposition is applied, providing a straightforward way to distribute computational workload. As ELM continues evolving at the kilometer-scale (km-scale), particularly with integrating lateral flow modeling, decomposition strategies must also account for the increased workload and data movement. This paper introduces a flexible user-defined domain decomposition framework, allowing users to customize domain partitioning based on application requirements. The impact of different decomposition strategies is evaluated across various applications concerning computation, communication, and I/O. Results demonstrate that while 1D partitioning yields superior I/O performance, k-nearest neighbors (KNN) clustering effectively reduces inter-process communication overhead. This study lays the groundwork for scalable partitioning in large-scale land surface simulations, enhancing next-generation Earth system modeling.
The ability of machine learning (ML) classification models to resist small, targeted input perturbations—known as adversarial attacks—is a key measure of their safety and reliability. We show that floating-point non associativity (FPNA) coupled with asynchronous parallel programming on GPUs is sufficient to result in misclassification, without any perturbation to the input. Additionally, we show that this misclassification is particularly significant for inputs close to the decision boundary and that standard adversarial robustness results may be overestimated up to 4.6 when not considering machine-level details. We first study a linear classifier, before focusing on standard Graph Neural Network (GNN) architectures and datasets used in robustness assessments. We develop a novel black-box attack using Bayesian optimization to discover external workloads that can change the instruction scheduling which bias the output of reductions on GPUs and reliably lead to misclassification. Motivated by these results, we present a new learnable permutation (LP) gradient-based approach to learning floating-point operation orderings that lead to misclassifications. The LP approach provides a worst-case estimate in a computationally efficient manner, avoiding the need to run identical experiments tens of thousands of times over a potentially large set of possible GPU states or architectures. Finally, using instrumentation-based testing, we investigate parallel reduction ordering across different GPU architectures under external background workloads, when utilizing multi-GPU virtualization, and when applying power capping. Our results demonstrate that parallel reduction ordering varies significantly across architectures under the first two conditions, substantially increasing the search space required to fully test the effects of this parallel scheduler-based vulnerability. These results and the methods developed here can help to include machine-level considerations into adversarial robustness assessments, which can make a difference in safety and mission critical applications.
Complex percutaneous interventions often require high radiation doses likely to produce skin radiation injuries. We assessed the methodology used to select patients with potential skin injuries in cardiac procedures and in need of clinical follow-up. We evaluated peak skin dose and clinical follow-up in a case of radiodermatitis produced during a total occlusion recanalization. This prospective study followed CIRSE and ACC/AHA/SCAI recommendations for patient radiation dose management in interventional procedures carried out in a university hospital with a workload of 4200 interventional cardiac procedures per year. Patient dose reports were automatically transferred to a central database. Patients exceeding trigger levels for air kerma area product (500 Gy cm{sup 2}) and cumulative skin dose (5 Gy) were counseled and underwent follow-up for early detection of skin injuries, with dermatologic support. The Ethical Committee and the Quality Assurance and Radiation Safety Committee approved the program. During 2010, a total of 13 patients (3.0/1,000 that year) received dose values exceeding trigger levels in the cardiovascular institute. Only one patient, who had undergone two consecutive procedures resulting in 970 Gy cm{sup 2} and 13.0 Gy as cumulative skin dose, showed signs of serious radiodermatitis that resolved in 3.7 months. The remaining patients did not manifest skin lesions during follow-up, and whenever patient examination was not feasible as part of the follow-up, neither patients nor families reported any skin injuries. Peak skin dose calculation and close clinical follow-up were feasible and appropriate, with a moderate additional workload for the staff and satisfaction for the patient.
In a cloud computing paradigm, energy efficient allocation of different virtualized ICT resources (servers, storage disks, and networks, and the like) is a complex problem due to the presence of heterogeneous application (e.g., content delivery networks, MapReduce, web applications, and the like) workloads having contentious allocation requirements in terms of ICT resource capacities (e.g., network bandwidth, processing speed, response time, etc.). Several recent papers have tried to address the issue of improving energy efficiency in allocating cloud resources to applications with varying degree of success. However, to the best of our knowledge there is no published literature on this subject that clearly articulates the research problem and provides research taxonomy for succinct classification of existing techniques. Hence, the main aim of this paper is to identify open challenges associated with energy efficient resource allocation. In this regard, the study, first, outlines the problem and existing hardware and software-based techniques available for this purpose. Furthermore, available techniques already presented in the literature are summarized based on the energy-efficient research dimension taxonomy. The advantages and disadvantages of the existing techniques are comprehensively analyzed against the proposed research dimension taxonomy namely: resource adaption policy, objective function, allocation method, allocation operation, and interoperability.
The explosive demand for artificial intelligence (AI) workloads has led to a significant increase in silicon area dedicated to lower-precision computations on recent high-performance computing hardware designs. However, mixed-precision capabilities, which can achieve performance improvements of up to 8x compared to double-precision in extreme compute-intensive workloads, remain largely untapped in most scientific applications. A growing number of efforts have shown that mixed-precision algorithmic innovations can deliver superior performance without sacrificing accuracy. These developments should prompt computational scientists to seriously consider whether their scientific modeling and simulation applications could benefit from the acceleration offered by new hardware and mixed-precision algorithms. In this survey, we (1) review progress across diverse scientific domains—fluid dynamics, weather and climate, quantum chemistry, and computational genomics—that have begun adopting mixed-precision strategies; (2) examine state-of-the-art algorithmic techniques such as iterative refinement, splitting and emulation schemes, and adaptive precision solvers; (3) assess their implications for accuracy, performance, and resource utilization; and (4) survey the emerging software ecosystem that enables mixed-precision methods at scale. We conclude with perspectives and recommendations on cross-cutting opportunities, domain-specific challenges, and the role of co-design between application scientists, numerical analysts, and computer scientists. Collectively, this survey underscores that mixed-precision numerics can reshape computational science by aligning algorithms with the evolving landscape of hardware capabilities.
There are numerous opportunities for data centers to participate in demand response programs considering their large energy capacities, flexible working environments and workloads, redundant design and operation, etc. As a type of demand response, frequency regulation requires fast response, and its potential is not fully explored by data centers yet. This paper proposes a synergistic control strategy for data center frequency regulation which uses both IT and cooling systems. It combines power management techniques at the server level with control of the chilled water supply temperature to track the regulation signal from the electrical market. A frequency regulation flexibility factor is also proposed to increase the IT capacity for frequency regulation. The performance of the control strategy is studied through numerical simulations using an equation based object-oriented Modelica platform designed for data centers. Simulation results show that with well-tuned control parameters, data centers can provide frequency regulation service in both regulation up and down. The performance of data centers in providing frequency regulation service is largely influenced by the regulation capacity bid, frequency regulation flexibility factor, workload condition, and cooling mode of the cooling system, and not significantly influenced by the time constant of chillers. In addition, compared with a server-only control strategy, the proposed synergistic control strategy can provide an extra regulation capacity of 3% of the design power when chillers are activated. Here, when chillers are deactivated, both strategies have a similar regulation capacity.
Scientific discoveries are significantly rely on the analysis of large volumes of data from observations and simulations of complex phenomena. Scientists compose the complex analyses as workflows and execute them on large-scale HPC systems. The workflow structures are in contrast with monolithic single simulations that have often been the primary use case on HPC systems. At the same time, new storage paradigms such as Burst Buffers are becoming available on HPC platforms. In this paper, we analyze the performance characteristics of a Burst Buffer and two representative scientific workflows with the aim of optimizing the usage of a Burst Buffer, extending our previous analyses (Daley et al., 2016). Our key contributions are (a) developing a performance analysis methodology pertinent to Burst Buffers, (b) improving the use of a Burst Buffer in workflows with bandwidth-sensitive and metadata-sensitive I/O workloads, (c) highlighting the key data management challenges when incorporating a Burst Buffer in the studied scientific workflows.
Scientific workflows provide an opportunity for declarative computational experiment design in an intuitive and efficient way. A distributed workflow is typically executed on a variety of resources, and it uses a variety of computational algorithms or tools to achieve the desired outcomes. Such a variety imposes additional complexity in scheduling these workflows on large scale computers. As computation becomes more distributed, insights into expected workload that a workflow presents become critical for effective resource allocation. In this paper, we present a modular framework that leverages Machine Learning for creating precise performance predictions of a workflow. The central idea is to partition a workflow in such a way that makes the task of forecasting each atomic unit manageable and gives us a way to combine the individual predictions efficiently. We recognize a combination of an executable and a specific physical resource as a single module. This gives us a handle to characterize workload and machine power as a single unit of prediction. Overall, our modular technique of creating atomic modules and deployment of longest-path approach to estimate workflow performance, allows the framework to adapt to highly complex nested directed acyclic workflows and scale to new scenarios, since it does not make assumptions of underlying workflow structure. We present performance estimation results of independent workflow modules executed on the XSEDE SDSC Comet cluster using various Machine Learning algorithms. The results provide insights into the behavior and effectiveness of different algorithms in the context of scientific workflow performance prediction.
Here we present a unique function-as-a-service (FaaS) dataset capturing the use of the Globus Compute (previously funcX) platform. Globus Compute implements a federated model via which users may deploy endpoints on arbitrary remote computers, from the edge to high performance computing (HPC) cluster, and they may then invoke Python functions on those endpoints via a reliable cloud -hosted service. The dataset covers 31 weeks and includes 2121472 task submissions from 252 users executed on 580 remote computing endpoints. It includes 277386 registered functions. We describe the dataset and various observations, some that are similar to other FaaS datasets, for example, that 74% of tasks run for less than 1 s, and some that are unique to Globus Compute, for example, that endpoints are used in different ways and that the majority of functions are related to scientific computing and machine learning. To the best of our knowledge, this dataset represents the first federated FaaS dataset that includes user workloads, distributed computing endpoints, and analysis of registered function bodies. We expect the dataset to be useful for researching FaaS architectures, workload modeling, container warming, and other distributed computing architectures.
Purpose: The widespread use of intensity modulated radiation therapy (IMRT) for cervical cancer has been limited by internal target and normal tissue motion. Such motion increases the risk of underdosing the target, especially as planning margins are reduced in an effort to reduce toxicity. This study explored 2 adaptive strategies to mitigate this risk and proposes a new, automated method that minimizes replanning workload. Methods and Materials: Thirty patients with cervical cancer participated in a prospective clinical study and underwent pretreatment and weekly magnetic resonance (MR) scans over a 5-week course of daily external beam radiation therapy. Target volumes and organs at risk (OARs) were contoured on each of the scans. Deformable image registration was used to model the accumulated dose (the real dose delivered to the target and OARs) for 2 adaptive replanning scenarios that assumed a very small PTV margin of only 3 mm to account for setup and internal interfractional motion: (1) a preprogrammed, anatomy-driven midtreatment replan (A-IMRT); and (2) a dosimetry-triggered replan driven by target dose accumulation over time (D-IMRT). Results: Across all 30 patients, clinically relevant target dose thresholds failed for 8 patients (27%) if 3-mm margins were used without replanning. A-IMRT failed in only 3 patients and also yielded an additional small reduction in OAR doses at the cost of 30 replans. D-IMRT assured adequate target coverage in all patients, with only 23 replans in 16 patients. Conclusions: A novel, dosimetry-triggered adaptive IMRT strategy for patients with cervical cancer can minimize the risk of target underdosing in the setting of very small margins and substantial interfractional motion while minimizing programmatic workload and cost.
Although explicit time integration schemes require small computational efforts per time step, their efficiency is severely restricted by their stability limits. Indeed, the multi-scale nature of some physical processes combined with highly unstructured meshes can lead some elements to impose a severely small stable time step for a global problem. Multirate methods offer a way to increase the global efficiency by gathering grid cells in appropriate groups under local stability conditions. These methods are well suited to the discontinuous Galerkin framework. The parallelization of the multirate strategy is challenging because grid cells have different workloads. The computational cost is different for each sub-time step depending on the elements involved and a classical partitioning strategy is not adequate any more. In this paper, we propose a solution that makes use of multi-constraint mesh partitioning. It tends to minimize the inter-processor communications, while ensuring that the workload is almost equally shared by every computer core at every stage of the algorithm. Particular attention is given to the simplicity of the parallel multirate algorithm while minimizing computational and communication overheads. Our implementation makes use of the MeTiS library for mesh partitioning and the Message Passing Interface for inter-processor communication. Performance analyses for two and three-dimensional practical applications confirm that multirate methods preserve important computational advantages of explicit methods up to a significant number of processors.
Non-volatile devices, such as SSDs, will be an integral part of the deepening storage hierarchy on large-scale HPC systems. These devices can be on the compute nodes as part of a distributed burst buffer service or they can be external. Wherever they are located in the hierarchy, one critical design issue is the SSD endurance under the write-heavy workloads, such as the checkpoint I/O for scientific applications. For these environments, it is widely assumed that checkpoint operations can occur once every 60 min and for each checkpoint step as much as half of the system memory can be written out. Unfortunately, for large-scale HPC applications, the burst buffer SSDs can be worn out much more quickly given the extensive amount of data written at every checkpoint step. One possible solution is to control the amount of data written by reducing the checkpoint frequency. However, a direct effect caused by reduced checkpoint frequency is the increased vulnerability window of system failures and therefore potentially wasted computation time, especially for large-scale compute jobs.
High-Performance Computing (HPC) platforms are required to solve the most diverse large-scale scientific problems in various research areas, such as biology, chemistry, physics, and health sciences. Researchers use a multitude of scientific softwares, which have different requirements. These include input and output operations, which directly impact performance due to the existing difference in processing and data access speeds. Thus, supercomputers must efficiently handle mixed workload when storing data from the applications. Understanding the set of applications and their performance running in a supercomputer is paramount to understanding the storage system's usage, pinpointing possible bottlenecks, and guiding optimization techniques. This research proposes a methodology and visualization tool to evaluate a supercomputer's data storage infrastructure's performance, taking into account the diverse workload and demands of the system over a long period of operation. As a study case, we focus on the Santos Dumont supercomputer, identifying inefficient usage, problematic performance factors, and providing guidelines on how to tackle those issues.
Quality assurance (QA) of volumetric modulated arc therapy (VMAT) increases the workload significantly. We compared the results from 4 verification methods to establish an efficient VMAT QA. Planning for VMAT treatments was carried out for 40 consecutive patients. Pretreatment verifications were carried out with ion chamber array Physikalish-Technische Werkstätten (PTW729), electronic portal dosimetry (EPID), ion chamber measurements, and independent dose calculation with Diamond program. 2D analyses were made using the gamma analysis (3 mm distance to agreement and 3% dose difference relative to maximum, 10% dose threshold). Average point dose difference calculated by Eclipse relative to ion chamber measurements and Diamond were 0.1%±0.9% and 0.6%±2.2%, respectively. Average pass rate for PTW729 was 99.2%±1.9% and 98.3%±1.3% for EPID. The total required time (linac occupancy time given in parentheses) for each QA method was: PTW729 43.5 minutes (26.5 minutes), EPID 14.5 minutes (2.5 minutes), ion chamber 34.5 minutes (26.5 minutes), and Diamond 12.0 minutes (0 minute). The results were consistent and allowed us to establish an optimized protocol, considering safety and accuracy as well as workload, consisting of 2 verification methods: EPID 2D analysis and independent dose calculation.