A Comparative Analysis of DAOS for Simulation and Deep Learning HPC Workloads
Explore the source record for details and available documents.
SEARCH · Engineering Papers
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.
Explore the source record for details and available documents.
Explore the source record for details and available documents.
Explore the source record for details and available documents.
Explore the source record for details and available documents.
Explore the source record for details and available documents.
Explore the source record for details and available documents.
High performance computing (HPC) scheduling landscape currently faces new challenges due to thechanges in the workload. Previously, HPC centers were dominated by tightly coupled MPI jobs. HPCworkloads increasingly include high-throughput, data-intensive, and stream-processing applications. Asa consequence, workloads are becoming more diverse at both application and job levels, posing newchallenges to classical HPC schedulers. There is a need to understand the current HPC workloads andtheir evolution to facilitate informed future scheduling research and enable efficient scheduling in futureHPC systems.In this paper, we present a methodology to characterize workloads and assess their heterogeneity,at a particular time period and its evolution over time. We apply this methodology to the workloads ofthree systems (Hopper, Edison, and Carver) at the National Energy Research Scientific Computing Center(NERSC). We present the resulting characterization of jobs, queues, heterogeneity, and performance thatincludes detailed information of a year of workload (2014) and evolution through the systems’ lifetime(2010–2014).
This paper presents a workload classification framework that discriminates illicit computation from authorized workloads on GPU-accelerated HPC systems. As such systems become more and more powerful, they are exploited by attackers to run malicious and for-profit programs that typically require extremely high computing ability to be successful. Our classification framework leverages the distinctive signatures between illicit and authorized workloads, and explore machine learning methods to learn the workloads and classify them. The framework uses lightweight, non-intrusive workload profiling to collect model input data, and explores multiple machine learning methods, particularly recurrent neural network (RNN) that is suitable for online anomalous workload detection. Evaluation results on four generations of GPU machines demonstrate that the workload classification framework can tell apart the illicit authorized workloads with a high accuracy of over 95%.
This paper presents a workload classification framework that discriminates illicit computation from authorized workloads on GPU-accelerated HPC systems. As such systems become more and more powerful, they are exploited by attackers to run malicious and for-profit programs that typically require extremely high computing ability to be successful. Our classification framework leverages the distinctive signatures between illicit and authorized workloads, and explore machine learning methods to learn the workloads and classify them. The framework uses lightweight, non-intrusive workload profiling to collect model input data, and explores multiple machine learning methods, particularly recurrent neural network (RNN) that is suitable for online anomalous workload detection. Evaluation results on three generations of GPU machines demonstrate that the workload classification framework can tell apart the illicit authorized workloads with a high accuracy of over 95%.
We explore an online reinforcement learning (RL) paradigm to dynamically optimize parallel particle tracing performance in distributed-memory systems. Our method combines three novel components: (1) a work donation algorithm, (2) a high-order workload estimation model, and (3) a communication cost model. First, we design an RL-based work donation algorithm. Our algorithm monitors workloads of processes and creates RL agents to donate data blocks and particles from high-workload processes to low-workload processes to minimize program execution time. The agents learn the donation strategy on the fly based on reward and cost functions designed to consider processes' workload changes and data transfer costs of donation actions. Second, we propose a workload estimation model, helping RL agents estimate the workload distribution of processes in future computations. Third, we design a communication cost model that considers both block and particle data exchange costs, helping RL agents make effective decisions with minimized communication costs. We demonstrate that our algorithm adapts to different flow behaviors in large-scale fluid dynamics, ocean, and weather simulation data. Our algorithm improves parallel particle tracing performance in terms of parallel efficiency, load balance, and costs of I/O and communication for evaluations with up to 16,384 processors.
Software simulators play a critical role in the development of new algorithms and system architectures in any field of engineering. Neuromorphic computing, which has shown potential in building brain-inspired energy-efficient hardware, suffers a slow-down in the development cycle due to a lack of flexible and easy-to-use simulators of either neuromorphic hardware itself or of spiking neural networks (SNNs), the type of neural network computation executed on most neuromorphic systems. While there are several openly available neuromorphic or SNN simulation packages developed by a variety of research groups, they have mostly targeted computational neuroscience simulations, and only a few have targeted small-scale machine learning tasks with SNNs. Evaluations or comparisons of these simulators have often targeted computational neuroscience-style workloads. In this work, we seek to evaluate the performance of several publicly available SNN simulators with respect to non-computational neuroscience workloads, in terms of speed, flexibility, and scalability. We evaluate the performance of the NEST, Brian2, Brian2GeNN, BindsNET and Nengo packages under a common front-end neuromorphic framework. Our evaluation tasks include a variety of different network architectures and workload types to mimic the computation common in different algorithms, including feed-forward network inference, genetic algorithms, and reservoir computing. We also study the scalability of each of these simulators when running on different computing hardware, from single core CPU workstations to multi-node supercomputers. Our results show that the BindsNET simulator has the best speed and scalability for most of the SNN workloads (sparse, dense, and layered SNN architectures) on a single core CPU. However, when comparing the simulators leveraging the GPU capabilities, Brian2GeNN outperforms the others for these workloads in terms of scalability. NEST performs the best for small sparse networks and is also the most flexible simulator in terms of reconfiguration capability NEST shows a speedup of at least 2x compared to the other packages when running evolutionary algorithms for SNNs. The multi-node and multi-thread capabilities of NEST show at least 2x speedup compared to the rest of the simulators (single core CPU or GPU based simulators) for large and sparse networks. We conclude our work by providing a set of recommendations on the suitability of employing these simulators for different tasks and scales of operations. We also present the characteristics for a future generic ideal SNN simulator for different neuromorphic computing workloads.
Traditional interconnection network design approaches focus on building general network topologies by optimizing the bisection bandwidth or minimizing the network’s diameter to reduce the maximum distance between any two nodes, thus amortizing the overall execution time of the HPC workloads. While such network topologies may accommodate a wide variety of applications in general, this may result in sub-optimal performance for many frequently-executed or dynamic workloads. In this paper, instead of focusing on designing an all-encompassing, general-purpose network topology, we develop a methodology to design customized network interconnects, evolved by “finding” the optimal topologies for a particular target workload given by its communication and contention profiles. To this end, we implement a Genetic Algorithm (GA)-based approach for network topology design tailored to improve the overall execution time of a particular workload of interest. We conducted extensive experiments with well-known motifs in physics-based workloads (Sweep3D and FFT), as well as with a representative graph application (MiniVite), using the well-known Structural Simulation Toolkit (SST) Macroscale Element Library (SST/macro) simulator for network interconnect evaluation. We demonstrate that our genetic algorithm-based approach is robust enough to find the underlying optimal topology of a particular workload.
Proportional to the scale increases in HPC systems, many scientific applications are becoming increasingly data intensive, and parallel I/O has become one of the dominant factors impacting the large-scale HPC application performance. On a typical large-scale HPC system, we have observed that the lack of a global workload coordination coupled with the shared nature of storage systems cause load imbalance and resource contention over the end-to-end I/O paths resulting in severe performance degradation. I/O load imbalance on HPC systems is generally a self-inflicted wound and mostly occurs between the I/O paths and resources consumed by each individual job. In this paper, we introduce TAPP-IO, a dynamic, shared load balancing framework for mitigating resource contention. TAPP-IO extends our previous work and solves two major limitations: First, it transparently intercepts file creation calls during runtime to balance the workload over all available storage targets. The usage of TAPP-IO requires no application source code modifications and is independent from any I/O middleware. The framework can be applied to almost any HPC platform and is suitable for systems that lack a centralized file system resource manager. Second, the framework proposes a new placement strategy to support not only file-per-process I/O, but also single shared file I/O. This opens the door to a new class of scientific applications that can leverage the placement library for improved performance. We demonstrate the effectiveness of our integration on the Titan system at the Oak Ridge National Laboratory. Our experiments with a synthetic benchmark and real-world HPC workload show that, even in a noisy production environment, TAPP-IO can improve large-scale application performance significantly.
Comprehensive analysis of high-performance computing (HPC) systems requires linking workload execution to system behavior. This kind of analysis is vital for diagnosing performance issues, managing capacity, detecting anomalous workloads, and understanding how applications interact with system hardware. This job-centric telemetry dataset unifies scheduler job records with node-level measurements, enabling direct association between workloads and their corresponding power, thermal, and performance characteristics. It contains sanitized, scheduler related metadata for 152,400 individual jobs that ran on the Frontier supercomputer and ended on selected days throughout 2024 and 2025, a subpopulation of ~6.8% of the total number of allocated jobs with non-zero run time on the system over that same period. Each is linked with files that contain telemetry time series records of the power utilization and temperature behavior of its allocated nodes and their processors during the run time of the job. Where available, a portion of the job files also contain network performance time series. Jobs are sampled from select days that reflect normal levels of user activity and possess job size distributions with large numbers of leadership class jobs (>20% of Frontier nodes). Jobs in this dataset attempt to best represent successful user workflows.
Deployment and execution of machine learning tasks on extreme-scale computing platforms face several significant technical challenges: 1) High computing cost incurred by dense networks – The computing workload of deep networks with densely-connected topology increases rapidly with the network size, imposing a non-scalable computing model of extreme-scale computing platforms; 2) Non-optimized workload distribution – Many advanced deep learning algorithms, e.g., sparsification and irregular net-work topology, produce very unbalanced workload distribution on extreme-scale computing platforms. The computation efficiency is greatly hindered by the incurred data and computation redundancies as well as long tails of the node with extensive workload; 3) Constraints in data movement and I/O bottle-neck – Inter-node data movement in extreme-scale computing platforms are associated with high energy and latency costs, and subject to the constraints of I/O bandwidth; and 4) Generalization of algorithm realization and acceleration on computing platforms – The large varieties of machine learning algorithms and structures of extreme-scale computing platforms make the derivation of a generalized algorithm realization and acceleration method very challenging, which, however, is the requirement by domain scientists and interested users. We call the above challenges Smale’s Problems in Machine Learning and Understanding for High-Performance Computing Scientific Discovery. The objective of our three-year research project is to develop a holistic innovation set at structure, assembly, and acceleration layers of machine learning algorithms to address the above challenges in algorithm deployment and execution. Three tasks are particularly performed, including: At the algorithm structure level, we investigate the techniques that can structurally sparsify on the topology of deep networks for computing workload reduction. We also study clustering and pruning techniques that can optimize the workload distributions over the extreme-scale computing platforms; At the algorithm assembly level, we derive a unified learning framework for unsupervised transfer learning and dynamic growing capabilities. Novel training methods are also exploited to enhance the training efficiency of the proposed framework; At the algorithm acceleration level, we will develop a series of techniques that can accelerate the computation of sparse matrix operations, which are one of the core executions in deep learning and optimize memory access of the concerned platforms. Our proposed techniques attack the fundamental problems in machine learning algorithms running on extreme-scale computing platforms by vertically integrating the solutions at three closely entangled layers, paving the long-term scaling path of machine learning applications under DOE context. Three tasks corresponding to the above respective research orientations are performed during the three-year project period with our collaborators at ORNL. The outcome of the proposed project is anticipated to form a holistic solution set of novel algorithms and network topologies, efficient training techniques, and fast acceleration methods to promote the computing scalability of the machine learning applications of particular interest to DOE.
Although solid-state drives (SSDs) offer significant performance improvements over hard disk drives (HDDs) for a number of workloads, they can exhibit substantial variance in request latency and throughput as a result of garbage collection (GC). When GC conflicts with an I/O stream, the stream can make no forward progress until the GC cycle completes. GC cycles are scheduled by logic internal to the SSD based on several factors such as the pattern, frequency, and volume of write requests. When SSDs are used in a RAID with currently available technology, the lack of coordination of the SSD-local GC cycles amplifies this performance variance. We propose a global garbage collection (GGC) mechanism to improve response times and reduce performance variability for a RAID of SSDs. We include a high-level design of SSD-aware RAID controller and GGC-capable SSD devices and algorithms to coordinate the GGC cycles. We develop reactive and proactive GC coordination algorithms and evaluate their I/O performance and block erase counts for various workloads. Our simulations show that GC coordination by a reactive scheme improves average response time and reduces performance variability for a wide variety of enterprise workloads. For bursty, write-dominated workloads, response time was improved by 69% and performance variability was reduced by 71%. We show that a proactive GC coordination algorithm can further improve the I/O response times by up to 9% and the performance variability by up to 15%. We also observe that it could increase the lifetimes of SSDs with some workloads (e.g. Financial) by reducing the number of block erase counts by up to 79% relative to a reactive algorithm for write-dominant enterprise workloads.
In modern multi-core processors, last-level caches (LLCs) are typically shared among multiple cores. Previous works have shown that such sharing is beneficial as different workloads have different needs for cache capacity, and logical partitioning of capacity can improve system performance. However, what is missing in previous works on partitioning shared LLCs is that the heterogeneity in spatial locality among workloads has not been explored. In other words, all the cores use the same block/line size in shared LLCs. In this work, we highlight that exploiting spatial locality enables much more effective cache sharing. The fundamental reason is that for many memory intensive workloads, their cache capacity requirements can be drastically reduced when a large block size is employed, therefore they can effectively donate more capacity to other workloads. To leverage spatial locality for cache partitioning effectively, we first propose a simple yet effective mechanism to measure both spatial and temporal locality at run-time. The locality information is then used to determine both the proper block size and the capacity assigned to each workload. Our experiments show that our Spatial Locality-aware Cache Partitioning (SLCP) significantly outperforms the previous works. We also present several case studies that dissect the effectiveness of SLCP compared to the existing approaches.
Multi-accelerator servers are increasingly being deployed in shared multi-tenant environments (such as in cloud data centers) in order to meet the demands of large-scale compute-intensive workloads. In addition, these accelerators are increasingly being inter-connected in complex topologies and workloads are exhibiting a wider variety of inter-accelerator communication patterns. However, existing al-location policies are ill-suited for these emerging use-cases. Specifically, this work identifies that multi-accelerator workloads are commonly fragmented leading to reduced bandwidth and increased latency for inter-accelerator communication. We propose Multi-Accelerator Pattern Allocation (MAPA), a graph pattern mining approach towards providing generalized allocation support for allocating multi-accelerator workloads on multi-accelerator servers. We demonstrate that MAPA is able to improve the execution time of multi-accelerator workloads and that MAPA is able to provide generalized benefits across various accelerator topologies. Finally, we demonstrate a speedup of 12.4% for75th percentile of jobs with the worst case execution time reduced by up to 35% against baseline policy using MAPA.