Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Multiple processors”

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.

At least 145 records · Page 8

Shared performance monitor in a multiprocessor system

A performance monitoring unit (PMU) and method for monitoring performance of events occurring in a multiprocessor system. The multiprocessor system comprises a plurality of processor devices units, each processor device for generating signals representing occurrences of events in the processor device, and, a single shared counter resource for performance monitoring. The performance monitor unit is shared by all processor cores in the multiprocessor system. The PMU comprises: a plurality of performance counters each for counting signals representing occurrences of events from one or more the plurality of processor units in the multiprocessor system; and, a plurality of input devices for receiving the event signals from one or more processor devices of the plurality of processor units, the plurality of input devices programmable to select event signals for receipt by one or more of the plurality of performance counters for counting, wherein the PMU is shared between multiple processing units, or within a group of processors in the multiprocessing system. The PMU is further programmed to monitor event signals issued from non-processor devices.

97 MATHEMATICS AND COMPUTING↗

Parallel Agent-Based Simulations on Clusters of GPUs and Multi-Core Processors

An effective latency-hiding mechanism is presented in the parallelization of agent-based model simulations (ABMS) with millions of agents. The mechanism is designed to accommodate the hierarchical organization as well as heterogeneity of current state-of-the-art parallel computing platforms. We use it to explore the computation vs. communication trade-off continuum available with the deep computational and memory hierarchies of extant platforms and present a novel analytical model of the tradeoff. We describe our implementation and report preliminary performance results on two distinct parallel platforms suitable for ABMS: CUDA threads on multiple, networked graphical processing units (GPUs), and pthreads on multi-core processors. Message Passing Interface (MPI) is used for inter-GPU as well as inter-socket communication on a cluster of multiple GPUs and multi-core processors. Results indicate the benefits of our latency-hiding scheme, delivering as much as over 100-fold improvement in runtime for certain benchmark ABMS application scenarios with several million agents. This speed improvement is obtained on our system that is already two to three orders of magnitude faster on one GPU than an equivalent CPU-based execution in a popular simulator in Java. Thus, the overall execution of our current work is over four orders of magnitude faster when executed on multiple GPUs.

97 MATHEMATICS AND COMPUTING↗

Thread-level parallelization and optimization of NWChem for the Intel MIC architecture

In the multicore era it was possible to exploit the increase in on-chip parallelism by simply running multiple MPI processes per chip. Unfortunately, manycore processors' greatly increased thread- and data-level parallelism coupled with a reduced memory capacity demand an altogether different approach. In this paper we explore augmenting two NWChem modules, triples correction of the CCSD(T) and Fock matrix construction, with OpenMP in order that they might run efficiently on future manycore architectures. As the next NERSC machine will be a self-hosted Intel MIC (Xeon Phi) based supercomputer, we leverage an existing MIC testbed at NERSC to evaluate our experiments. In order to proxy the fact that future MIC machines will not have a host processor, we run all of our experiments in native mode. We found that while straightforward application of OpenMP to the deep loop nests associated with the tensor contractions of CCSD(T) was sufficient in attaining high performance, significant e ort was required to safely and efeciently thread the TEXAS integral package when constructing the Fock matrix. Ultimately, our new MPI+OpenMP hybrid implementations attain up to 65× better performance for the triples part of the CCSD(T) due in large part to the fact that the limited on-card memory limits the existing MPI implementation to a single process per card. Additionally, we obtain up to 1.6× better performance on Fock matrix constructions when compared with the best MPI implementations running multiple processes per card.

CCSD(T)↗

Thread-Level Parallelization and Optimization of NWChem for the Intel MIC Architecture

In the multicore era it was possible to exploit the increase in on-chip parallelism by simply running multiple MPI processes per chip. Unfortunately, manycore processors' greatly increased thread- and data-level parallelism coupled with a reduced memory capacity demand an altogether different approach. In this paper we explore augmenting two NWChem modules, triples correction of the CCSD(T) and Fock matrix construction, with OpenMP in order that they might run efficiently on future manycore architectures. As the next NERSC machine will be a self-hosted Intel MIC (Xeon Phi) based supercomputer, we leverage an existing MIC testbed at NERSC to evaluate our experiments. In order to proxy the fact that future MIC machines will not have a host processor, we run all of our experiments in tt native mode. We found that while straightforward application of OpenMP to the deep loop nests associated with the tensor contractions of CCSD(T) was sufficient in attaining high performance, significant effort was required to safely and efficiently thread the TEXAS integral package when constructing the Fock matrix. Ultimately, our new MPI OpenMP hybrid implementations attain up to 65x better performance for the triples part of the CCSD(T) due in large part to the fact that the limited on-card memory limits the existing MPI implementation to a single process per card. Additionally, we obtain up to 1.6x better performance on Fock matrix constructions when compared with the best MPI implementations running multiple processes per card.

97 MATHEMATICS AND COMPUTING↗

AthenaK: A Performance-portable Version of the Athena++ Adaptive Mesh Refinement Framework

We describe AthenaK: a new implementation of the Athena++ block-based adaptive mesh refinement framework using the Kokkos programming model. Finite volume methods for Newtonian, special relativistic, and general relativistic (GR) hydrodynamics and magnetohydrodynamics (MHD), and GR-radiation hydrodynamics and MHD, as well as a module for evolving Lagrangian tracer or charged test particles (e.g., cosmic rays) are implemented using the framework. In two companion papers, we describe (1) a new solver for the Einstein equations based on the Z4c formalism, and (2) a GRMHD solver in dynamical spacetimes also implemented using the framework, enabling new applications in numerical relativity. By adopting Kokkos, the code can be run on virtually any hardware, including CPUs, GPUs from multiple vendors, and emerging Advanced RISC Machine processors. AthenaK shows excellent performance and weak scaling, achieving over 1 billion cell updates per second for hydrodynamics in three dimensions on a single NVIDIA Grace Hopper processor. It does this with a typical parallel efficiency of 80% on 65,536 AMD GPUs on the OLCF Frontier system. Such performance portability enables AthenaK to leverage modern exascale computing systems for challenging applications in astrophysical fluid dynamics, numerical relativity, and multimessenger astrophysics.

79 ASTRONOMY AND ASTROPHYSICS↗

Generation-based memory synchronization in a multiprocessor system with weakly consistent memory accesses

In a multiprocessor system, a central memory synchronization module coordinates memory synchronization requests responsive to memory access requests in flight, a generation counter, and a reclaim pointer. The central module communicates via point-to-point communication. The module includes a global OR reduce tree for each memory access requesting device, for detecting memory access requests in flight. An interface unit is implemented associated with each processor requesting synchronization. The interface unit includes multiple generation completion detectors. The generation count and reclaim pointer do not pass one another.

97 MATHEMATICS AND COMPUTING↗

Generation-based memory synchronization in a multiprocessor system with weakly consistent memory accesses

In a multiprocessor system, a central memory synchronization module coordinates memory synchronization requests responsive to memory access requests in flight, a generation counter, and a reclaim pointer. The central module communicates via point-to-point communication. The module includes a global OR reduce tree for each memory access requesting device, for detecting memory access requests in flight. An interface unit is implemented associated with each processor requesting synchronization. The interface unit includes multiple generation completion detectors. The generation count and reclaim pointer do not pass one another.

97 MATHEMATICS AND COMPUTING↗

Compressing micro-operations in scheduler entries in a processor

An electronic device includes a processor having a micro-operation queue, multiple scheduler entries, and scheduler compression logic. When a pair of micro-operations in the micro-operation queue is compressible in accordance with one or more compressibility rules, the scheduler compression logic acquires the pair of micro-operations from the micro-operation queue and stores information from both micro-operations of the pair of micro-operations into different portions in a single scheduler entry. In this way, the scheduler compression logic compresses the pair of micro-operations into the single scheduler entry.

Boyer, Michael W.↗

Neutron multiplicity counting: Confidence intervals for reconstruction parameters

From nuclear materials accountability to homeland security, the need for improved nuclear material detection, assay, and authentication has grown over the past decades. Starting in the 1940s, neutron multiplicity counting techniques have enabled quantitative evaluation of masses and multiplications of fissile materials. In this paper, we propose a new method to compute uncertainties on these parameters using a model-based sequential Bayesian processor, resulting in credible regions in the fissile material mass and multiplication space. These uncertainties will enable us to evaluate quantitatively proposed improvements to the theoretical fission chain model. Additionally, because the processor can calculate uncertainties in real time, it is a useful tool in applications such as portal monitoring: monitoring can stop as soon as a preset confidence of non-threat is reached.

11 NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗

Dynamic load balancing of matrix-vector multiplications on roadrunner compute nodes

Hybrid architectures that combine general purpose processors with accelerators are being adopted in several large-scale systems such as the petaflop Roadrunner supercomputer at Los Alamos. In this system, dual-core Opteron host processors are tightly coupled with PowerXCell 8i processors within each compute node. In this kind of hybrid architecture, an accelerated mode of operation is typically used to offload performance hotspots in the computation to the accelerators. In this paper we explore the suitability of a variant of this acceleration mode in which the performance hotspots are actually shared between the host and the accelerators. To achieve this we have designed a new load balancing algorithm, which is optimized for the Roadrunner compute nodes, to dynamically distribute computation and associated data between the host and the accelerators at runtime. Results are presented using this approach for sparse and dense matrix-vector multiplications that show load-balancing can improve performance by up to 24% over solely using the accelerators.

97 MATHEMATICS AND COMPUTING↗

In-Situ Statistical Analysis of Autotune Simulation Data using Graphical Processing Units

Developing accurate building energy simulation models to assist energy efficiency at speed and scale is one of the research goals of the Whole-Building and Community Integration group, which is a part of Building Technologies Research and Integration Center (BTRIC) at Oak Ridge National Laboratory (ORNL). The aim of the Autotune project is to speed up the automated calibration of building energy models to match measured utility or sensor data. The workflow of this project takes input parameters and runs EnergyPlus simulations on Oak Ridge Leadership Computing Facility’s (OLCF) computing resources such as Titan, the world’s second fastest supercomputer. Multiple simulations run in parallel on nodes having 16 processors each and a Graphics Processing Unit (GPU). Each node produces a 5.7 GB output file comprising 256 files from 64 simulations. Four types of output data covering monthly, daily, hourly, and 15- minute time steps for each annual simulation is produced. A total of 270TB+ of data has been produced. In this project, the simulation data is statistically analyzed in-situ using GPUs while annual simulations are being computed on the traditional processors. Titan, with its recent addition of 18,688 Compute Unified Device Architecture (CUDA) capable NVIDIA GPUs, has greatly extended its capability for massively parallel data processing. CUDA is used along with C/MPI to calculate statistical metrics such as sum, mean, variance, and standard deviation leveraging GPU acceleration. The workflow developed in this project produces statistical summaries of the data which reduces by multiple orders of magnitude the time and amount of data that needs to be stored. These statistical capabilities are anticipated to be useful for sensitivity analysis of EnergyPlus simulations

accelerators↗

System and method for memory allocation in a multiclass memory system

A system for memory allocation in a multiclass memory system includes a processor coupleable to a plurality of memories sharing a unified memory address space, and a library store to store a library of software functions. The processor identifies a type of a data structure in response to a memory allocation function call to the library for allocating memory to the data structure. Using the library, the processor allocates portions of the data structure among multiple memories of the multiclass memory system based on the type of the data structure.

97 MATHEMATICS AND COMPUTING↗

Adaptive track scheduling to optimize concurrency and vectorization in GeantV

The GeantV project is focused on the R&D of new particle transport techniques to maximize parallelism on multiple levels, profiting from the use of both SIMD instructions and co-processors for the CPU-intensive calculations specific to this type of applications. In our approach, vectors of tracks belonging to multiple events and matching different locality criteria must be gathered and dispatched to algorithms having vector signatures. While the transport propagates tracks and changes their individual states, data locality becomes harder to maintain. The scheduling policy has to be changed to maintain efficient vectors while keeping an optimal level of concurrency. The model has complex dynamics requiring tuning the thresholds to switch between the normal regime and special modes, i.e. prioritizing events to allow flushing memory, adding new events in the transport pipeline to boost locality, dynamically adjusting the particle vector size or switching between vector to single track mode when vectorization causes only overhead. Lastly, this work requires a comprehensive study for optimizing these parameters to make the behaviour of the scheduler self-adapting, presenting here its initial results.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

Power system model parameter conditioning tool

A power system model parameter conditioning tool including a server control processor in communication with phasor measurement unit monitored data records of multiple disturbance events, a model calibration unit providing event screening, power system model simulation, and simultaneous tuning of model parameters. The model calibration performing a simulation using default model parameters, the processor comparing the simulation results to the monitored data. If the prediction is within threshold, then terminating conditioning; else performing parameter identifiability analysis to determine differing effects of various model parameters on power system model accuracy, selecting a parameter set causing a degradation in power system model prediction, and updating the default model parameters corresponding to members of the parameter set with values selected to reduce the degradation. A method and a non-transitory computer readable medium are also disclosed.

Baone, Chaitanya Ashok↗

Sparse Matrix-Matrix Multiplication on Multilevel Memory Architectures: Algorithms and Experiments

Architectures with multiple classes of memory media are becoming a common part of mainstream supercomputer deployments. So called multi-level memories offer differing characteristics for each memory component including variation in bandwidth, latency and capacity. This paper investigates the performance of sparse matrix multiplication kernels on two leading highperformance computing architectures — Intel's Knights Landing processor and NVIDIA's Pascal GPU. We describe a data placement method and a chunking-based algorithm for our kernels that exploits the existence of the multiple memory spaces in each hardware platform. We evaluate the performance of these methods w.r.t. standard algorithms using the auto-caching mechanisms Our results show that standard algorithms that exploit cache reuse performed as well as multi-memory-aware algorithms for architectures such as Ki\iLs where the memory subsystems have similar latencies. However, for architectures such as GPUS where memory subsystems differ significantly in both bandwidth and latency, multi-memory-aware methods are crucial for good performance. In addition, our new approaches permit the user to run problems that require larger capacities than the fastest memory of each compute node without depending on the software-managed cache mechanisms.

97 MATHEMATICS AND COMPUTING↗

Efficient Sorting on the Tilera Manycore Architecture

e present an efficient implementation of the radix sort algo- rithm for the Tilera TILEPro64 processor. The TILEPro64 is one of the first successful commercial manycore processors. It is com- posed of 64 tiles interconnected through multiple fast Networks- on-chip and features a fully coherent, shared distributed cache. The architecture has a large degree of flexibility, and allows various optimization strategies. We describe how we mapped the algorithm to this architecture. We present an in-depth analysis of the optimizations for each phase of the algorithm with respect to the processor’s sustained performance. We discuss the overall throughput reached by our radix sort implementation (up to 132 MK/s) and show that it provides comparable or better performance-per-watt with respect to state-of-the art implemen- tations on x86 processors and graphic processing units.

manycore↗

The ASC Sequoia Programming Model

In the late 1980's and early 1990's, Lawrence Livermore National Laboratory was deeply engrossed in determining the next generation programming model for the Integrated Design Codes (IDC) beyond vectorization for the Cray 1s series of computers. The vector model, developed in mid 1970's first for the CDC 7600 and later extended from stack based vector operation to memory to memory operations for the Cray 1s, lasted approximately 20 years (See Slide 5). The Cray vector era was deemed an extremely long lived era as it allowed vector codes to be developed over time (the Cray 1s were faster in scalar mode than the CDC 7600) with vector unit utilization increasing incrementally over time. The other attributes of the Cray vector era at LLNL were that we developed, supported and maintained the Operating System (LTSS and later NLTSS), communications protocols (LINCS), Compilers (Civic Fortran77 and Model), operating system tools (e.g., batch system, job control scripting, loaders, debuggers, editors, graphics utilities, you name it) and math and highly machine optimized libraries (e.g., SLATEC, and STACKLIB). Although LTSS was adopted by Cray for early system generations, they later developed COS and UNICOS operating systems and environment on their own. In the late 1970s and early 1980s two trends appeared that made the Cray vector programming model (described above including both the hardware and system software aspects) seem potentially dated and slated for major revision. These trends were the appearance of low cost CMOS microprocessors and their attendant, departmental and mini-computers and later workstations and personal computers. With the wide spread adoption of Unix in the early 1980s, it appeared that LLNL (and the other DOE Labs) would be left out of the mainstream of computing without a rapid transition to these 'Killer Micros' and modern OS and tools environments. The other interesting advance in the period is that systems were being developed with multiple 'cores' in them and called Symmetric Multi-Processor or Shared Memory Processor (SMP) systems. The parallel revolution had begun. The Laboratory started a small 'parallel processing project' in 1983 to study the new technology and its application to scientific computing with four people: Tim Axelrod, Pete Eltgroth, Paul Dubois and Mark Seager. Two years later, Eugene Brooks joined the team. This team focused on Unix and 'killer micro' SMPs. Indeed, Eugene Brooks was credited with coming up with the 'Killer Micro' term. After several generations of SMP platforms (e.g., Sequent Balance 8000 with 8 33MHz MC32032s, Allian FX8 with 8 MC68020 and FPGA based Vector Units and finally the BB&N Butterfly with 128 cores), it became apparent to us that the killer micro revolution would indeed take over Crays and that we definitely needed a new programming and systems model. The model developed by Mark Seager and Dale Nielsen focused on both the system aspects (Slide 3) and the code development aspects (Slide 4). Although now succinctly captured in two attached slides, at the time there was tremendous ferment in the research community as to what parallel programming model would emerge, dominate and survive. In addition, we wanted a model that would provide portability between platforms of a single generation but also longevity over multiple--and hopefully--many generations. Only after we developed the 'Livermore Model' and worked it out in considerable detail did it become obvious that what we came up with was the right approach. In a nutshell, the applications programming model of the Livermore Model posited that SMP parallelism would ultimately not scale indefinitely and one would have to bite the bullet and implement MPI parallelism within the Integrated Design Code (IDC). We also had a major emphasis on doing everything in a completely standards based, portable methodology with POSIX/Unix as the target environment. We decided against specialized libraries like STACKLIB for performance, but kept as many general purpose, portable math libraries as were needed by the codes. Third, we assumed that the SMPs in clusters would evolve in time to become more powerful, feature rich and, in particular, offer more cores. Thus, we focused on OpenMP, and POSIX PThreads for programming SMP parallelism. These code porting efforts were lead by Dale Nielsen, A-Division code group leader, and Randy Christensen, B-Division code group leader. Most of the porting effort revolved removing 'Crayisms' in the codes: artifacts of LTSS/NLTSS, Civic compiler extensions beyond Fortran77, IO libraries and dealing with new code control languages (we switched to Perl and later to Python). Adding MPI to the codes was initially problematic and error prone because the programmers used MPI directly and sprinkled the calls throughout the code.

99 GENERAL AND MISCELLANEOUS↗

Broadband unidirectional visible imaging using wafer-scale nano-fabrication of multi-layer diffractive optical processors

We present a broadband and polarization-insensitive unidirectional imager that operates at the visible part of the spectrum, where image formation occurs in one direction, while in the opposite direction, it is blocked. This approach is enabled by deep learning-driven diffractive optical design with wafer-scale nano-fabrication using high-purity fused silica to ensure optical transparency and thermal stability. Our design achieves unidirectional imaging across three visible wavelengths (covering red, green, and blue parts of the spectrum), and we experimentally validated this broadband unidirectional imager by creating high-fidelity images in the forward direction and generating weak, distorted output patterns in the backward direction, in alignment with our numerical simulations. This work demonstrates wafer-scale production of diffractive optical processors, featuring 16 levels of nanoscale phase features distributed across two axially aligned diffractive layers for visible unidirectional imaging. This approach facilitates mass-scale production of ~0.5 billion nanoscale phase features per wafer, supporting high-throughput manufacturing of hundreds to thousands of multi-layer diffractive processors suitable for large apertures and parallel processing of multiple tasks. Beyond broadband unidirectional imaging in the visible spectrum, this study establishes a pathway for artificial-intelligence-enabled diffractive optics with versatile applications, signaling a new era in optical device functionality with industrial-level, massively scalable fabrication.

36 MATERIALS SCIENCE↗