On the Memory Underutilization: Exploring Disaggregated Memory on HPC Systems
Explore the source record for details and available documents.
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.
Explore the source record for details and available documents.
Summary Disaggregated memory is a promising approach that addresses the limitations of traditional memory architectures by enabling memory to be decoupled from compute nodes and shared across a data center. Cloud platforms have deployed such systems to improve overall system memory utilization, but performance can vary across workloads. High‐performance computing (HPC) is crucial in scientific and engineering applications, where HPC machines also face the issue of underutilized memory. As a result, improving system memory utilization while understanding workload performance is essential for HPC operators. Therefore, learning the potential of a disaggregated memory system before deployment is a critical step. This paper proposes a methodology for exploring the design space of a disaggregated memory system. It incorporates key metrics that affect performance on disaggregated memory systems: memory capacity, local and remote memory access ratio, injection bandwidth, and bisection bandwidth, providing an intuitive approach to guide machine configurations based on technology trends and workload characteristics. We apply our methodology to analyze thirteen diverse workloads, including AI training, data analysis, genomics, protein, fusion, atomic nuclei, and traditional HPC bookends. Our methodology demonstrates the ability to comprehend the potential and pitfalls of a disaggregated memory system and provides motivation for machine configurations. Our results show that eleven of our thirteen applications can leverage injection bandwidth disaggregated memory without affecting performance, while one pays a rack bisection bandwidth penalty and two pay the system‐wide bisection bandwidth penalty. In addition, we also show that intra‐rack memory disaggregation would meet the application's memory requirement and provide enough remote memory bandwidth.
Many modern applications have memory footprints that are increasingly large, driving system memory capacities higher and higher. Moreover, these systems are often organized where the bulk of the memory is collocated with the compute capability, which necessitates the need for message passing APIs to facilitate information sharing between compute nodes. Due to the diversity of applications that must run on High-Performance Computing (HPC) systems, the memory utilization can fluctuate wildly from one application to another. And, because memory is located in the node, maintenance can become problematic because each node must be taken offline and upgraded individually. To address these issues, vendors are exploring disaggregated, memory-centric, systems. In this type of organization, there are discrete nodes, reserved solely for memory, which are shared across many compute nodes. Due to their capacity, low-power, and non-volatility, Non-Volatile Memories (NVMs) are ideal candidates for these memory nodes. This report discusses a new component for the Structural Simulation Toolkit (SST), Opal, that can be used to study the impact of using NVMs in a disaggregated system in terms of performance, security, and memory management. This page intentionally left blank.
Recent advances in integrated photonics enable the implementation of reconfigurable, high-bandwidth, and low energy-per-bit interconnects in next-generation data centers. We propose and evaluate an Optically Connected Memory (OCM) architecture that disaggregates the main memory from the computation nodes in data centers. OCM is based on micro-ring resonators (MRRs), and it does not require any modification to the DRAM memory modules. We calculate energy consumption from real photonic devices and integrate them into a system simulator to evaluate performance. Here, our results show that (1) OCM is capable of interconnecting four DDR4 memory channels to a computing node using two fibers with 1.02 pJ energy-per-bit consumption and (2) OCM performs up to 5.5× faster than a disaggregated memory with 40G PCIe NIC connectors to computing nodes.
Compute Express Link (CXL) is an important emerging standard for disaggregated memory. While this standard provisions coherency across numerous hosts and devices, implementing hardware support for type three devices is challenging. In this work, we look at the overhead of software synchronization and using software-based coherency. Moreover, we discuss the limits of software-based coherency in fully expressing modern synchronization techniques for a CXL-based disaggregate memory system. We demonstrate our approach using a CXL hardware prototype and running a version of the famous Peterson Lock (enhanced to run with more than two threads). We analyze its performance and share how more advanced synchronization techniques might interact with software-based coherence CXL hardware and program execution models.
Abstract not provided.
Abstract not provided.
Abstract not provided.
Abstract not provided.
To provide affordable energy-saving solutions for the small and medium-sized manufacturers (SMMs), we propose a unified framework for generating predictive models that support real-time disaggregation of power consumption from combined inputs, enabling automatic machine state identification simultaneously for joint analysis of energy usage patterns. Further, the proposed framework transforms raw power consumption into a time series with look-back and bootstrap capabilities for historical pattern detection, while a learning architecture utilizes the stacked long short-term memory (LSTM) layers as encoders for embedding generation with sequential awareness. Experimental results demonstrate 93.65% minimum accuracy in ideal case of real-time energy usage and machine state prediction.
The expected halt of traditional technology scaling is motivating increased heterogeneity in high-performance computing (HPC) systems with the emergence of numerous specialized accelerators. As heterogeneity increases, so does the risk of underutilizing expensive hardware resources if we preserve today’s rigid node configuration and reservation strategies. This has sparked interest in resource disaggregation to enable finer-grain allocation of hardware resources to applications. However, there is currently no data-driven study of what range of disaggregation is appropriate in HPC. To that end, we perform a detailed analysis of key metrics sampled in NERSC’s Cori, a production HPC system that executes a diverse open-science HPC workload. In addition, we profile a variety of deep-learning applications to represent an emerging workload. We show that for a rack (cabinet) configuration and applications similar to Cori, a central processing unit with intra-rack disaggregation has a 99.5% probability to find all resources it requires inside its rack. In addition, ideal intra-rack resource disaggregation in Cori could reduce memory and NIC resources by 5.36% to 69.01% and still satisfy the worst-case average rack utilization.
In this report, we examine several emerging technologies of interest to the Department of Energy and its computational centers. These include: 1) quantifying the benefit of tighter CPU-GPU integration, 2) quantifying the appropriate CPU core:GPU ratio, 3) quantifying the penalty for CPU-GPU disaggregation, 4) quantifying the benefits of tighter GPU-GPU integration, 5) quantifying the benefits of unified memory, and 6) quantifying the benefits of tighter FPGA-GPU integration.
Matrix-vector multiplication is the key operation in any Krylov-subspace iteration method. We are interested in Krylov methods applied to problems associated with the graph Laplacian arising from large scale-free graphs. Furthermore, computations with graphs of this type on parallel distributed-memory computers are challenging. This is due to the fact that scale-free graphs have a degree distribution that follows a power law, and currently available graph partitioners are not efficient for such an irregular degree distribution. The lack of a good partitioning leads to excessive interprocessor communication requirements during every matrix-vector product. Here, we present an approach to alleviate this problem based on embedding the original irregular graph into a more regular one by disaggregating (splitting up) vertices in the original graph. The matrix-vector operations for the original graph are performed via a factored triple matrix-vector product involving the embedding graph. And even though the latter graph is larger, we are able to decrease the communication requirements considerably and improve the performance of the matrix-vector product.
Here, we discuss the development of an agent-based modelling software development kit, and the implementation and validation of a model using it that integrates dynamic simulation of travel demand, network supply and network operations. A description is given of the core utilities in the kit: a parallel discrete event engine, interprocess exchange engine, and memory allocator, as well as a number of ancillary utilities: visualization library, database IO library, and scenario manager. The overall framework emphasizes the design goals of: generality, code agility, and high performance. This framework allows the modeling of several aspects of transportation system that are typically done with separate stand-alone software applications, in a high-performance and extensible manner. The issue of integrating such models as dynamic traffic assignment and disaggregate demand models has been a long standing issue for transportation modelers. The integrated approach shows a possible way to resolve this difficulty. The simulation model built from the POLARIS framework is a single, shared-memory process for handling all aspects of the integrated urban simulation. The resulting gains in computational efficiency and performance allow planning models to be extended to include previously separate aspects of the urban system, enhancing the utility of such models from the planning perspective. Lastly, initial tests with case studies involving traffic management center impacts on various network events such as accidents, congestion and weather events, show the potential of the system.
The problem of noninstrusive load monitoring (NILM) is usually formulated as a single-channel blind source separation task, whose successful solution enable fast and convenient load identification and energy disaggregation. When applied at test time, NILM algorithms aim to identify the operating characteristics of individual appliances from an aggregate power measurement of the entire house. Recent advances in deep learning gave rise to many methods that mostly focus on learning a direct mapping from aggregate measurement to individual appliance power. However, these methods are not only computationally expensive, but they often suffer from overfitting and do not generalize very well. In this article, we propose a novel NILM method that leverages advances in statistical learning that have not been properly applied in this domain before. The proposed method consists of three stages: first, a Bayesian nonparametric learning-based approach for appliance state extraction; second, synthetic minority oversampling technique for data augmentation and mitigating the heavy imbalance in switching events; and third, appliance-specific lightweight long short-term memory networks for status classification for each appliance. Here, we adopt a “differential” input (the difference before and after the switching event) to reduce the complexity of network training and make the proposed method robust to multiappliance switching events. Experiments are conducted to demonstrate the effectiveness of the proposed method, achieving superior performance when compared to recent methods. An ablation study is conducted to demonstrate the effectiveness of each module of our method. Finally, we investigate the quality of generated synthetic samples.