Evaluating the Benefits of An Extended Memory Hierarchy for Parallel Streamline Algorithms
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.
The LBNL Visualization research program focuses on a key set of problems facing our community as we move towards the exascale regime of computing. First, we aim to better understand how to effectively take advantage of an evolving architectural landscape where there are an increasing number of cores per processor by examining algorithmic reformulation and alternative methods of programming these increasingly complex platforms. Second, as the gap widens between computational and I/O capacity, it is increasingly imperative to perform more visual data exploration and analysis operations concurrent with the simulation to avoid increasingly expensive I/O. Our approach is to examine several interrelated issues aimed at minimizing the memory footprint, data movement, and design/implementation issues related to deploying tools that can have broad applicability across many different simulation codes and science domains. Third, increasing computational power produces larger and more complex scientific data, and our research program focuses on new technologies aimed at enabling scientific data understanding with ever-larger data. Here, we explore how multi-resolution data representation techniques can reduce data movement, and study novel approaches that focus visualization and analysis processing on subsets of data that are “scientifically interesting,” thereby also minimizing data movement while accelerating scientific knowledge discovery.
As the memory and storage hierarchy get deeper and more complex, it is important to have new benchmarks and evaluation tools that allow us to explore the emerging middleware solutions to use this hierarchy. Skel is a tool aimed at automating and refining this process of studying HPC I/O performance. It works by generating application I/O kernel/benchmarks as determined by a domain-specific model. This paper provides some techniques for extending Skel to address new situations and to answer new research questions. For example, we document use cases as diverse as using Skel to troubleshoot I/O performance issues for remote users, refining an I/O system model, and facilitating the development and testing of a mechanism for runtime monitoring and performance analytics. We also discuss data oriented extensions to Skel to support the study of compression techniques for Exascale scientific data management.
Significance The generalized Langevin equation (GLE) provides a precise description of coarse-grained variable dynamics in reduced dimension models. However, computation of the memory kernel poses a major challenge to the practical use of the GLE. This paper presents a data-driven approach to compute the memory kernel, relying on a hierarchy of parameterized rational approximations in terms of the Laplace transform, which can be expanded to arbitrarily high order as necessary. This parameterization makes it convenient to represent the GLE via an extended stochastic model where the memory term is eliminated by properly introducing auxiliary variables. The present method is well-suited for constructing reduced models for nonequilibrium properties of complex systems such as biomolecules, chemical reaction networks, and climate simulations.
Tensor algebra operations represent an important class of algorithms used across many applications, including machine learning, scientific computing, and data analytics. As a result, the efficient generation of custom accelerators for tensor operations has received increased attention. Previous efforts have produced automated tools enabling users to prototype and explore optimized accelerators. However, little effort has been focused on the host-accelerator interaction in these tools. Efficient use of hardware accelerators requires knowledge about the accelerator's capabilities (operations, data formats, and opcode support), the host CPU microarchitecture (e.g., memory hierarchy), the host-accelerator interface, and the application's features (which code regions should be mapped onto an accelerator). Manually rewriting the original applications to facilitate improved custom accelerator mapping is an error-prone and time-consuming endeavor. To cope with this, we propose AXI4MLIR, a new framework to automatically generate and optimize the communication between the host CPU and arbitrary accelerators that implement linear algebra algorithms. AXI4MLIR extends the MLIR compiler framework to automatically generate efficient host-accelerator driver code for accelerators with AXI-based interfaces. Our compiler extensions enable automatic driver code generation while carefully considering the host's memory hierarchy and target accelerator features. To demonstrate the flexibility and utility of AXI4MLIR, we test it with diverse use cases that include different types of accelerators, tiling scenarios, and dataflow schemes. We compare our experimental results to manual implementations of host-accelerator driver code and find that our approach can reduce CPU cache references by 56% and deliver up to a 1.65x speedup.
Performance modeling is a challenging problem due to the complexities of hardware architectures. In this paper, we present PPT-GPU, a scalable and accurate simulation framework that enables GPU code developers and architects to predict the performance of applications in a fast, and accurate manner on different GPU architectures. PPT-GPU is part of the open source project, Performance Prediction Toolkit (PPT) developed at the Los Alamos National Laboratory. We extend the old GPU model in PPT that predict the runtimes of computational physics codes to offer better prediction accuracy, for which, we add models for different memory hierarchies found in GPUs and latencies for different instructions. To further show the utility of PPT-GPU, we compare our model against real GPU device(s) and the widely used cycle-accurate simulator, GPGPU-Sim using different workloads from RODINIA and Parboil benchmarks. The results indicate that the predicted performance of PPT-GPU is within a 10% error compared to the real device(s). In addition, PPT-GPU is highly scalable, where it is up to 450x faster than GPGPU-Sim with more accurate results.
A system and method of supporting point-to-point synchronization among processes/nodes implementing different hardware barriers in a tuple space/coordinated namespace (CNS) extended memory storage architecture. The system-wide CNS provides an efficient means for storing data, communications, and coordination within applications and workflows implementing barriers in a multi-tier, multi-nodal tree hierarchy. The system provides a hardware accelerated mechanism to support barriers between the participating processes. Also architected is a tree structure for a barrier processing method where processes are mapped to nodes of a tree, e.g., a tree of degree k to provide an efficient way of scaling the number of processes in a tuple space/coordination namespace.
A system and method of supporting point-to-point synchronization among processes/nodes implementing different hardware barriers in a tuple space/coordinated namespace (CNS) extended memory storage architecture. The system-wide CNS provides an efficient means for storing data, communications, and coordination within applications and workflows implementing barriers in a multi-tier, multi-nodal tree hierarchy. The system provides a hardware accelerated mechanism to support barriers between the participating processes. Also architected is a tree structure for a barrier processing method where processes are mapped to nodes of a tree, e.g., a tree of degree k, to provide an efficient way of scaling the number of processes in a tuple space/coordination namespace.
Checkpoint restart plays an important role in high-performance computing (HPC) applications, allowing simulation runtime to extend beyond a single job allocation and facilitating recovery from hardware failure. Yet, as machines grow in size and in complexity, traditional approaches to checkpoint restart are becoming prohibitive. Current methods store a subset of the application’s state and exploit the memory hierarchy in the machine. However, as the energy cost of data movement continues to dominate, further reductions in checkpoint size are needed. Lossy compression, which can significantly reduce checkpoint sizes, offers a potential to reduce computational cost in checkpoint restart. This article investigates the use of numerical properties of partial differential equation (PDE) simulations, such as bounds on the truncation error, to evaluate the feasibility of using lossy compression in checkpointing PDE simulations. Restart from a checkpoint with lossy compression is considered for a fail-stop error in two time-dependent HPC application codes: PlasComCM and Nek5000. Results show that error in application variables due to a restart from a lossy compressed checkpoint can be masked by the numerical error in the discretization, leading to increased efficiency in checkpoint restart without influencing overall accuracy in the simulation.
Exascale systems are expected to feature hundreds of thousands of compute nodes with hundreds of hardware threads and complex memory hierarchies with a mix of on-package and persistent memory modules. In this context, the Argo project is developing a new operating system for exascale machines. Targeting production workloads using workflows or coupled codes, we improve the Linux kernel on several fronts. We extend the memory management of Linux to be able to subdivide NUMA memory nodes, allowing better resource partitioning among processes running on the same node. We also add support for memory-mapped access to node- local, PCIe-attached NVRAM devices and introduce a new scheduling class targeted at parallel runtimes supporting user- level load balancing.
Computer architecture experts expect that non-volatile memory (NVM) hierarchies will play a more significant role in future systems including mobile, enterprise, and HPC architectures. With this expectation in mind, we present NVL-C: a novel programming system that facilitates the efficient and correct programming of NVM main memory systems. The NVL-C programming abstraction extends C with a small set of intuitive language features that target NVM main memory, and can be combined directly with traditional C memory model features for DRAM. We have designed these new features to enable compiler analyses and run-time checks that can improve performance and guard against a number of subtle programming errors, which, when left uncorrected, can corrupt NVM-stored data. Moreover, to enable recovery of data across application or system failures, these NVL-C features include a flexible directive for specifying NVM transactions. So that our implementation might be extended to other compiler front ends and languages, the majority of our compiler analyses are implemented in an extended version of LLVM's intermediate representation (LLVM IR). We evaluate NVL-C on a number of applications to show its flexibility, performance, and correctness.
In order to add more capabilities to Halide, we have designed a new framework called Tiramisu and integrated this framework into Halide. Since Tiramisu enables Halide to target heterogeneous architectures, our development efforts have been refocused on Tiramisu. Most high-performance computer systems today are complex and increasingly heterogeneous; they may have CPUs, GPUs and FPGAs. Achieving best performance requires taking full advantage of all these different architectures. To address this issue, we have designed Tiramisu, an optimization framework that enables Halide (and other DSLs) to target heterogeneous architectures. Tiramisu is an optimization framework that takes as input a high level, architecture-independent representation of code and a set of scheduling and data mapping commands that guide code transformation. The input can either be generated by a domain-specific language (DSL) compiler such as Halide or directly written by a programmer. Tiramisu then applies the user-specified code and data-layout transformations and generates an architecture-specific, low-level intermediate representation (IR) that takes advantage of modern architectural features such as multicore parallelism, non-uniform memory (NUMA) hierarchies, clusters, and accelerators like GPUs and FPGAs. We integrated Tiramisu within Halide and implemented a representative set of benchmarks to evaluate this integration. Tiramisu is now open source and is available for public use (http://tiramisu-compiler.org/). A paper about Tiramisu was published, it shows that Tiramisu extends Halide with many new capabilities and that Tiramisu can generate efficient code for multicores, GPUs, FPGAs and distributed heterogeneous systems. The performance of code generated by the Tiramisu backends matches or exceeds hand optimized reference implementations. For example, the multicore backend matches the highly optimized Intel MKL library on many kernels and shows speedups reaching 4x over the original Halide. In addition to making Tiramisu more robust, we have used Tiramisu to implement a set of representative tensor operation for constructing baryon building blocks required for multi baryon contractions in LQCD. In order to implement this code, we needed to generalize Tiramisu in two ways: first we needed to support indirect array accesses, and second, we needed to add support for complex numbers to Tiramisu. The code generated by Tiramisu is 6x faster than the reference code. Our efforts towards an MPI based multi-node version of tiramisu have matured and the resulting code scales well on multiple nodes (tests up to 512 KNL nodes have been undertaken).
Multitime quantum correlation functions are central objects in physical science, offering a direct link between the experimental observables and the dynamics of an underlying model. While experiments such as 2D spectroscopy and quantum control can now measure such quantities, the accurate simulation of such responses remains computationally expensive and sometimes impossible, depending on the system’s complexity. A natural tool to employ is the generalized quantum master equation (GQME), which can offer computational savings by extending reference dynamics at a comparatively trivial cost. However, dynamical methods that can tackle chemical systems with atomistic resolution, such as those in the semiclassical hierarchy, often suffer from poor accuracy, limiting the credence one might lend to their results. By combining work on the accuracy-boosting formulation of semiclassical memory kernels with recent work on the multitime GQME, here we show for the first time that one can exploit a multitime semiclassical GQME to dramatically improve both the accuracy of coarse mean-field Ehrenfest dynamics and obtain orders of magnitude efficiency gains.
The NASA Multiscale Analysis Tool (NASMAT) is a platform for multiscale modeling of composites which can perform analysis of materials with any arbitrary number of length scales. The platform supports modularity, scalability, and interoperability using recursive procedures and data structures. A Macro solver driven parallelization scheme often limits the capability of NASMAT to scale as it has access to limited memory and number of cores (often one core/thread) and often forces to implement macro solver specific changes to the platform. In this work, a partitioned task-parallel approach is adopted, where the parallelization strategy adopted for NASMAT is independent of the macro solver and the computational resources are managed independently. The programming architecture takes into account the hierarchy of multiple scales (task-dependence) and the heterogeneous nature (dynamic load balancing) of computation through implementation of a hierarchy-informed task parallel model. The partitioned nature of the framework further extends the “plug and play” capability of NASMAT. preCICE, an open-source library for coupling multiphysics solver in a partitioned manner, is adopted to integrate NASMAT with an external macro solver by implementing a NASMAT adapter for preCICE. Speedup and scalability of the framework is studied for micromechanical models of varying size.
Over the past decade, high performance computing has evolved rapidly, not only in hardware architectures but also with increasing complexity of real applications. Technologies have been developing to aim at scaling up to thousands of processors on both distributed and shared memory systems. Development of parallel programs on these computers is always a challenging task. Today, writing parallel programs with message passing (e.g. MPI) is the most popular way of achieving scalability and high performance. However, writing message passing programs is difficult and error prone. Recent years new effort has been made in defining new parallel programming paradigms. The best examples are: HPF (based on data parallelism) and OpenMP (based on shared memory parallelism). Both provide simple and clear extensions to sequential programs, thus greatly simplify the tedious tasks encountered in writing message passing programs. HPF is independent of memory hierarchy, however, due to the immaturity of compiler technology its performance is still questionable. Although use of parallel compiler directives is not new, OpenMP offers a portable solution in the shared-memory domain. Another important development involves the tremendous progress in the internet and its associated technology. Although still in its infancy, Java promisses portability in a heterogeneous environment and offers possibility to "compile once and run anywhere." In light of testing these new technologies, we implemented new parallel versions of the NAS Parallel Benchmarks (NPBs) with HPF and OpenMP directives, and extended the work with Java and Java-threads. The purpose of this study is to examine the effectiveness of alternative programming paradigms. NPBs consist of five kernels and three simulated applications that mimic the computation and data movement of large scale computational fluid dynamics (CFD) applications. We started with the serial version included in NPB2.3. Optimization of memory and cache usage was applied to several benchmarks, noticeably BT and SP, resulting in better sequential performance. In order to overcome the lack of an HPF performance model and guide the development of the HPF codes, we employed an empirical performance model for several primitives found in the benchmarks. We encountered a few limitations of HPF, such as lack of supporting the "REDISTRIBUTION" directive and no easy way to handle irregular computation. The parallelization with OpenMP directives was done at the outer-most loop level to achieve the largest granularity. The performance of six HPF and OpenMP benchmarks is compared with their MPI counterparts for the Class-A problem size in the figure in next page. These results were obtained on an SGI Origin2000 (195MHz) with MIPSpro-f77 compiler 7.2.1 for OpenMP and MPI codes and PGI pghpf-2.4.3 compiler with MPI interface for HPF programs.
Graph problems are common across fields of scientific computing and social sciences. However, despite their importance, implementing graph algorithms effectively on modern computing systems is a challenging task that requires significant programming effort and generally results in customized implementations. Current computing and memory hierarchies are not architected for irregular computations resulting in challenges for graph algorithms to achieve high performance on those architectures. In this paper, we present GraphX, a novel compiler framework and DSL designed to simplify the development of efficient graph algorithms and achieve high performance on modern computing systems. GraphX consists of a DSL for efficient implementation of graph algorithms, various optimizations, such as support for sparse linear algebra and workspace transformations, optimized graph primitives, including semiring and masking, and a high-performance code generation engine. Using GraphX, users can implement graph algorithms using a semantically-rich language with graph-oriented operators. GraphX uses these semantics to automatically generate efficient code for target architectures, increasing performance and portability across architectures. The composable nature of GraphX makes it possible to extend the set of optimizations and architectures without modifying the source code. We demonstrate GraphX outperforms state-of-the-art graph libraries, such as LAGraph, up to $3.7 speedup in semiring operations, $2.19 speedup in an important sparse computational kernel, and $9.05 speedup in graph processing algorithms.
This report summarizes the current statistical analysis capability of OVIS and how it works in conjunction with the OVIS data readers and interpolators. It also documents how to extend these capabilities. OVIS is a tool for parallel statistical analysis of sensor data to improve system reliability. Parallelism is achieved using a distributed data model: many sensors on similar components (metaphorically sheep) insert measurements into a series of databases on computers reserved for analyzing the measurements (metaphorically shepherds). Each shepherd node then processes the sheep data stored locally and the results are aggregated across all shepherds. OVIS uses the Visualization Tool Kit (VTK) statistics algorithm class hierarchy to perform analysis of each process's data but avoids VTK's model aggregation stage which uses the Message Passing Interface (MPI); this is because if a single process in an MPI job fails, the entire job will fail. Instead, OVIS uses asynchronous database replication to aggregate statistical models. OVIS has several additional features beyond those present in VTK that, first, accommodate its particular data format and, second, improve the memory and speed of the statistical analyses. First, because many statistical algorithms are multivariate in nature and sensor data is typically univariate, interpolation of data is required to provide simultaneous observations of metrics. Note that in this report, we will refer to a single value obtained from a sensor as a measurement while a collection of multiple sensor values simultaneously present in the system is an observation. A base class for interpolation is provided that abstracts the operation of converting multiple sensor measurements into simultaneous observations. A concrete implementation is provided that performs piecewise constant temporal interpolation of multiple metrics across a single component. Secondly, because calculations may summarize data too large to fit in memory OVIS analyses batches of observations at a time and aggregates these intermediate intra-process models as it goes before storing the final model for inter-process aggregation via database replication. This reduces the memory footprint of the analysis, interpolation, and the database client and server query processing. This also interleaves processing with the disk I/O required to fetch data from the database - also improving speed. This report documents how OVIS performs analyses and how to create additional analysis components that fetch measurements from the database, perform interpolation, or perform operations on streamed observations (such as model updates or assessments). The rest of this section outlines the OVIS analysis algorithm and is followed by sections specific to each subtask. Note that we are limiting our discussion for now to the creation of a model from a set of measurements, and not including the assessment of observations using a model. The same framework can be used for assessment but that use case is not detailed in this report.
The pre-exascale systems will soon be deployed with a deep, complex memory hierarchy composed of many heterogeneous memories. This presents multiple challenges for users including: how to allocate data objects with locality between memories and devices for the various memories in these systems, which includes DRAM, High-bandwidth Memory (HBM), and non-volatile random access memory (NVRAM), and how to perform these allocations while providing portability for their application. Currently, the user can make use of multiple, disjoint libraries to allocate data objects on these memories. However, it is difficult to obtain locality between memories and devices when using libraries that are unaware of each other. This paper presents the Unified Memory Allocator (UMA) of the SHARed data-structure centric Programming abstraction (SharP) library, which provides a unified interface for memory allocations across DRAM, HBM, and NVRAM and is extensible to support future memory types. In addition, the SharP UMA allows for portability between systems by supporting both explicit and implicit, intent-based memory allocations. To demonstrate the ease of use of the SharP UMA, we have extended both Open MPIand OpenSHMEM-Xto support SharP. We validate this work by evaluating the performance implications and intent-based approach with synthetic benchmarks as well as adaptations of the Graph500 benchmark.