Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “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 1,441 records · Page 80

Utilizing the ISS Mission as a Testbed to Develop Cognitive Communications Systems

The ISS provides an excellent opportunity for pioneering artificial intelligence software to meet the challenges of real-time communications (comm) link management. This opportunity empowers the ISS Program to forge a testbed for developing cognitive communications systems for the benefit of the ISS mission, manned Low Earth Orbit (LEO) science programs and future planetary exploration programs. In November, 1998, the Flight Operations Directorate (FOD) started the ISS Antenna Manager (IAM) project to develop a single processor supporting multiple comm satellite tracking for two different antenna systems. Further, the processor was developed to be highly adaptable as it supported the ISS mission through all assembly stages. The ISS mission mandated communications specialists with complete knowledge of when the ISS was about to lose or gain comm link service. The current specialty mandated cognizance of large sun-tracking solar arrays and thermal management panels in addition to the highly-dynamic satellite service schedules and rise/set tables. This mission requirement makes the ISS the ideal communications management analogue for future LEO space station and long-duration planetary exploration missions. Future missions, with their precision-pointed, dynamic, laser-based comm links, require complete autonomy for managing high-data rate communications systems. Development of cognitive communications management systems that permit any crew member or payload science specialist, regardless of experience level, to control communications is one of the greater benefits the ISS can offer new space exploration programs. The IAM project met a new mission requirement never previously levied against US space-born communications systems management: process and display the orientation of large solar arrays and thermal control panels based on real-time joint angle telemetry. However, IAM leaves the actual communications availability assessment to human judgement, which introduces unwanted variability because each specialist has a different core of experience with comm link performance. Because the ISS utilizes two different frequency bands, dynamic structure can be occasionally translucent at one frequency while it can completely interdict service at the other frequency. The impact of articulating structure on the comm link can depend on its orientation at the time it impinges on the link. It can become easy for a human specialist to cross-associate experience at one frequency with experience at the other frequency. Additionally, the specialist's experience is incremental, occurring one nine-hour shift at a time. Only the IAM processor experiences the complete 24x7x365 communications link performance for both communications links but, it has no "learning capability." If the IAM processor could be endowed with a cognitive ability to remember past structure-induced comm link outages, based on its knowledge of the ISS position, attitude, communications gear, array joint angles and tracking accuracy, it could convey such experience to the human operator. It could also use its learned communications link behaviors to accurately convey the availability of future communications sessions. Further, the tool could remember how accurately or inaccurately it predicted availability and correct future predictions based on past performance. The IAM tool could learn frequency-specific impacts due to spacecraft structures and pass that information along as "experience." Such development would provide a single artificial intelligence processor that could provide two different experience bases. If it also "knew" the satellite service schedule, it could distinguish structure blockage from schedule or planet blockage and then quickly switch to another satellite. Alternatively, just as a human operator could judge, a cognizant comm system based on the IAM model could "know" that the blockage is not going to last very long and continue tracking a comm satellite, waiting for it to track away from structure. Ultimately, once this capability was fully developed and tested in the Mission Control Center, it could be transferred on-orbit to support development of operations concepts that include more advanced cognitive communications systems. Future applications of this capability are easily foreseen because even more dynamic satellite constellations with more nodes and greater capability are coming. Currently, the ISS fully employs its high-data-rate return link for harvesting payload science. In the coming months, it will double that data rate and is forecast to fully utilize that capability. Already there is talk of an upgrade that quadruples the current data rate allocated to ISS payload science before the end of its mission and laser comm links have already been tested from the ISS. Every data rate upgrade mandates more complicated and sensitive communications equipment which implies greater expertise invested in the human operator. Future on-orbit cognizant comm systems will be needed to meet greater performance demands aboard larger, far more complicated spacecraft. In the LEO environment, the old-style one-satellite-per-spacecraft operations concept will give way to a new concept of a single customer spacecraft simultaneously using multiple comm satellites. Much more highly-dynamic manned LEO missions with decades of crew members potentially increase the demand for communications link performance. A cognizant on-board communications system will meet advanced communications demands from future LEO missions and future planetary missions. The ISS has fledgling components of future exploration programs, both LEO and planetary. Further, the Flight Operations Directorate, through the IAM project, has already begun to develop a communications management system that attempts to solve advanced problems ideally represented by dynamic structure impacting scheduled satellite service. With an earnest project to integrate artificial intelligence into the IAM processor, the ISS Program could develop a cognizant communications system that could be adapted and transferred to future on-orbit avionics designs.

Jackson, Dan↗

Digital Beamforming Scatterometer

This paper discusses scatterometer measurements collected with multi-mode Digital Beamforming Synthetic Aperture Radar (DBSAR) during the SMAP-VEX 2008 campaign. The 2008 SMAP Validation Experiment was conducted to address a number of specific questions related to the soil moisture retrieval algorithms. SMAP-VEX 2008 consisted on a series of aircraft-based.flights conducted on the Eastern Shore of Maryland and Delaware in the fall of 2008. Several other instruments participated in the campaign including the Passive Active L-Band System (PALS), the Marshall Airborne Polarimetric Imaging Radiometer (MAPIR), and the Global Positioning System Reflectometer (GPSR). This campaign was the first SMAP Validation Experiment. DBSAR is a multimode radar system developed at NASA/Goddard Space Flight Center that combines state-of-the-art radar technologies, on-board processing, and advances in signal processing techniques in order to enable new remote sensing capabilities applicable to Earth science and planetary applications [l]. The instrument can be configured to operate in scatterometer, Synthetic Aperture Radar (SAR), or altimeter mode. The system builds upon the L-band Imaging Scatterometer (LIS) developed as part of the RadSTAR program. The radar is a phased array system designed to fly on the NASA P3 aircraft. The instrument consists of a programmable waveform generator, eight transmit/receive (T/R) channels, a microstrip antenna, and a reconfigurable data acquisition and processor system. Each transmit channel incorporates a digital attenuator, and digital phase shifter that enables amplitude and phase modulation on transmit. The attenuators, phase shifters, and calibration switches are digitally controlled by the radar control card (RCC) on a pulse by pulse basis. The antenna is a corporate fed microstrip patch-array centered at 1.26 GHz with a 20 MHz bandwidth. Although only one feed is used with the present configuration, a provision was made for separate corporate feeds for vertical and horizontal polarization. System upgrades to dual polarization are currently under way. The DBSAR processor is a reconfigurable data acquisition and processor system capable of real-time, high-speed data processing. DBSAR uses an FPGA-based architecture to implement digitally down-conversion, in-phase and quadrature (I/Q) demodulation, and subsequent radar specific algorithms. The core of the processor board consists of an analog-to-digital (AID) section, three Altera Stratix field programmable gate arrays (FPGAs), an ARM microcontroller, several memory devices, and an Ethernet interface. The processor also interfaces with a navigation board consisting of a GPS and a MEMS gyro. The processor has been configured to operate in scatterometer, Synthetic Aperture Radar (SAR), and altimeter modes. All the modes are based on digital beamforming which is a digital process that generates the far-field beam patterns at various scan angles from voltages sampled in the antenna array. This technique allows steering the received beam and controlling its beam-width and side-lobe. Several beamforming techniques can be implemented each characterized by unique strengths and weaknesses, and each applicable to different measurement scenarios. In Scatterometer mode, the radar is capable to.generate a wide beam or scan a narrow beam on transmit, and to steer the received beam on processing while controlling its beamwidth and side-lobe level. Table I lists some important radar characteristics

Rincon, Rafael F.↗

Methods and Apparatus for Aggregation of Multiple Pulse Code Modulation Channels into a Signal Time Division Multiplexing Stream

Methods and apparatus are provided for telemetry processing using a telemetry processor. The telemetry processor can include a plurality of communications interfaces, a computer processor, and data storage. The telemetry processor can buffer sensor data by: receiving a frame of sensor data using a first communications interface and clock data using a second communications interface, receiving an end of frame signal using a third communications interface, and storing the received frame of sensor data in the data storage. After buffering the sensor data, the telemetry processor can generate an encapsulated data packet including a single encapsulated data packet header, the buffered sensor data, and identifiers identifying telemetry devices that provided the sensor data. A format of the encapsulated data packet can comply with a Consultative Committee for Space Data Systems (CCSDS) standard. The telemetry processor can send the encapsulated data packet using a fourth and a fifth communications interfaces.

Chang, Chen J.↗

New Rover Conops with High-Performance Onboard Computing: Give Up Raw Data to Reduce Ops Cost and Do More Science

A major portion of time during the tactical operation of Mars rovers is spent for selecting, prioritizing, and coordinating sciences and engineering activities such that they fit within resource constraints, including the downlink data volume, energy, and time. In particular, the downlink data volume constraint is getting particularly tighter in recent missions because modern instruments produce increasingly high data volume while the communication bandwidth is essentially bounded by the law of physics. Tactical operation would be substantially simplified, hence the operation cost could be reduced, if the data volume constraint is relaxed or even removed. In this abstract, we propose a new operation paradigm for achieving this goal. The key observation is that, both in science and engineering applications, the bit size of raw data is typically much greater than the volume of processed information that is needed for scientific or engineering analysis. For example, a full-resolution image from Mastcam-Z, the main science camera on Perseverance, is about 700 kB in volume and we downlinked 29,685 images up to Sol 243, totaling ~20 GB of data. But of course, scientists do not use every pixel of these images; what they really look for in the images are geological features, typically represented by specific geometric configurations or textures. An end product after processing hundreds of Mascam-Z images could be a single geological map summarizing the spatial distribution of the features. For another example, a 100-meter drive of Perseverance produces 7-12 MB of drive telemetry, which records every detail of the rover's motion at 8 Hz, including position, attitude, steering angles, encoder readings, motor currents and many other information. But what the ground engineers eventually pay attention to is the signs of anomaly, such as excessive motor currents or high slip; if a drive is nominal, the vast majority of this data is unused. What if, then, we process the raw data onboard and only downlink the processed data that is relevant to scientific or engineering analyses, such as a list of detected science features (with cropped images) or a list of potential signs of anomaly while driving? A major roadblock for such onboard, high-level information processing has been the onboard computational resource. RAD750, the main onboard computer of Perseverance, is obviously not sufficient for performing complex image or signal processing such as object detection, semantic segmentation, or anomaly detection. Interestingly, RAD750 is not the best processor that Perseverance has; Qualcomm's Snapdragon 801, a modern mobile processor, is on her Heli Base Station, a device for communicating with Mars Helicopter Ingenuity; also, Intel's Atom E3845 processors are on engineering cameras. In the reminder of this paper, we will introduce two particular uses cases of these high-performance co-processors (meaning auxiliary CPU, GPU, or other types of processors that are separate from the main processor that runs the main flight software) for lowering operation cost and accommodating more science activities for a given communication constraint.

Didier, A.↗

Low latency memory access and synchronization

A low latency memory system access is provided in association with a weakly-ordered multiprocessor system. Bach processor in the multiprocessor shares resources, and each shared resource has an associated lock within a locking device that provides support for synchronization between the multiple processors in the multiprocessor and the orderly sharing of the resources. A processor only has permission to access a resource when it owns the lock associated with that resource, and an attempt by a processor to own a lock requires only a single load operation, rather than a traditional atomic load followed by store, such that the processor only performs a read operation and the hardware locking device performs a subsequent write operation rather than the processor. A simple prefetching for non-contiguous data structures is also disclosed. A memory line is redefined so that in addition to the normal physical memory data, every line includes a pointer that is large enough to point to any other line in the memory, wherein the pointers to determine which memory line to prefetch rather than some other predictive algorithm. This enables hardware to effectively prefetch memory access patterns that are non-contiguous, but repetitive.

Blumrich, Matthias A↗

Pausing and activating thread state upon pin assertion by external logic monitoring polling loop exit time condition

A system and method for enhancing performance of a computer which includes a computer system including a data storage device. The computer system includes a program stored in the data storage device and steps of the program are executed by a processer. The processor processes instructions from the program. A wait state in the processor waits for receiving specified data. A thread in the processor has a pause state wherein the processor waits for specified data. A pin in the processor initiates a return to an active state from the pause state for the thread. A logic circuit is external to the processor, and the logic circuit is configured to detect a specified condition. The pin initiates a return to the active state of the thread when the specified condition is detected using the logic circuit.

97 MATHEMATICS AND COMPUTING↗

On-package multiprocessor ground-referenced single-ended interconnect

A system of interconnected chips comprising a multi-chip module (MCM) includes a first processor chip, a second processor chip, and an MCM package configured to include the first processor chip, the second processor chip, and an interconnect circuit. The first processor chip is configured to include a first ground-referenced single-ended signaling (GRS) interface circuit. A first set of electrical traces fabricated within the MCM package and configured to couple the first GRS interface circuit to the interconnect circuit. The second processor chip is configured to include a second GRS interface circuit. A second set of electrical traces fabricated within the MCM package and configured to coupled the second GRS interface circuit to the interconnect circuit.

Dally, William J.↗

Data shuffling with hierarchical tuple spaces

Methods and systems for shuffling data are described. A processor may generate pair data from source data. The processor may insert the pair data into local tuple spaces. In response to a request for a particular key, the processor may determine a presence of the requested key in a global tuple space. The processor may, in response to a presence of the requested key in the global tuple space, update the global tuple space. The update may be based on the pair data among the local tuple spaces including the existing key. The processor may, in response to an absence of the requested key in the global tuple space, insert pair data including the missing key from the local tuple spaces into the global tuple space. The processor may fetch the requested pair data, and may shuffle the fetched data to generate a dataset.

Andrade Costa, Carlos Henrique↗

Optimization control technology for building energy conservation

A simulation processor generates and stores a simulation model based on conditions associated with a physical structure, such as a building. A neural network processor implements a neural network, having an input layer coupled to receive sensor data from the structure and having an output layer coupled to supply control signals to the at least one electrically operable environmental control device. The neural network is trained using the simulation model. A particle swarm optimization processor programmed to receive the simulation results and perform particle swarm optimization, ascertains optimal parameters for controlling the at least one electrically operable environmental control device and supplies these optimal parameters to the neural network processor. The neural network processor uses the optimal parameters supplied by the particle swarm optimization processor to further train the neural network.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗

Parallel 3-D method of characteristics in MPACT

A new parallel 3-D MOC kernel has been developed and implemented in MPACT which makes use of the modular ray tracing technique to reduce computational requirements and to facilitate parallel decomposition. The parallel model makes use of both distributed and shared memory parallelism which are implemented with the MPI and OpenMP standards, respectively. The kernel is capable of parallel decomposition of problems in space, angle, and by characteristic rays up to 0(104) processors. Initial verification of the parallel 3-D MOC kernel was performed using the Takeda 3-D transport benchmark problems. The eigenvalues computed by MPACT are within the statistical uncertainty of the benchmark reference and agree well with the averages of other participants. The MPACT k{sub eff} differs from the benchmark results for rodded and un-rodded cases by 11 and -40 pcm, respectively. The calculations were performed for various numbers of processors and parallel decompositions up to 15625 processors; all producing the same result at convergence. The parallel efficiency of the worst case was 60%, while very good efficiency (>95%) was observed for cases using 500 processors. The overall run time for the 500 processor case was 231 seconds and 19 seconds for the case with 15625 processors. Ongoing work is focused on developing theoretical performance models and the implementation of acceleration techniques to minimize the number of iterations to converge. (authors)

97 MATHEMATICS AND COMPUTING↗

Scalable load balancing for massively parallel distributed Monte Carlo particle transport

In order to run computer simulations efficiently on massively parallel computers with hundreds of thousands or millions of processors, care must be taken that the calculation is load balanced across the processors. Examining the workload of every processor leads to an unscalable algorithm, with run time at least as large as O(N), where N is the number of processors. We present a scalable load balancing algorithm, with run time 0(log(N)), that involves iterated processor-pair-wise balancing steps, ultimately leading to a globally balanced workload. We demonstrate scalability of the algorithm up to 2 million processors on the Sequoia supercomputer at Lawrence Livermore National Laboratory. (authors)

97 MATHEMATICS AND COMPUTING↗

Multi-node and Multi-core Performance Studies of a Monte Carlo code RMC

Over the past 15 years the architecture of high-performance computing systems has become increasingly heterogeneous with individual computing cores being subdivided into nodes and even auxiliary co-processing units such as graphics processing units or the Intel many integrated core co-processors. In fact, future computer platforms is moving toward a larger number of nodes and cores with lower memory available per node. Today, several production Monte Carlo codes use some types of hybrid parallelism strategy that makes use of shared-memory parallelism within a computing node and message-passing parallelism between nodes, which enables the codes to utilize available computing resources as much as possible, both among and within computing nodes on a cluster. The so-called hybrid parallelization is also possible with Reactor Monte Carlo Code RMC by activating both MPI and OpenMP parallel algorithms together. On HPC' which typically consists of several or several-ten thousand nodes equipping 16-32 computing intra-cores per node, simulation of RMC can be carried out in the hybrid parallelization with the MPI inter-node parallelization and the OpenMP intra-node parallelization. RMC has adopted a hybrid message-passing and shared-memory approach to parallelism, which enables the code to utilize available computing resources, both among and within computing nodes on a cluster. Shared-memory parallelism is especially valuable for large-model simulations because it enables multiple compute cores within a node to share common model information such as the geometry description, material, cross section and tally data, thus reducing the memory usage per processor. The calculation on the Tianhe-2 supercomputer took 6144 cores, resulting in a rate of 173,314,000 neutrons/minute for BEAVRS benchmark. While pin-wise full core H-M benchmark burnup calculations with millions of depletion regions was selected using MPI/OpenMP parallelism, the speed of 6,144 cores reach 1,843,200 burnable regions per minute. Keeping up with the continuing trend for increases in the number of computing resources within modern HPC systems is a significant challenge for teams that develop and maintain high-performance MC radiation transport solvers. order to address this challenge, RMC has adopted a hybrid message-passing and shared-memory approach to parallelism, which enables the code to utilize available computing resources, both among and within computing nodes on a cluster. Shared-memory parallelism is especially valuable for large-model simulations because it enables multiple compute cores within a node to share common model information such as the geometry description, material and composition definitions, and cross section data, thus reducing the memory usage per processor. The parallel performance has been studied by using two complex and representative benchmarks, H-M and BEAVRS benchmark. All simulations were run on Intel Xeon E5-2692 12 core, 2.2 GHz processors at MilkyWay-2 Super Computer with up to 12,288 available computing cores. While employing MPI parallelism, excellent parallel efficiency is achieved even above 12,000 processors and RMC reaches the parallel efficiency of 96% on 12,288 cores for the pin-by-pin computation of the BEAVRS benchmark. While BEAVRS criticality benchmark was selected using MPI/OpenMP parallelism, the speed of 6,144 cores reach 173,314,000 particles per minute. While pin-wise full core H-M benchmark burnup calculations with millions of depletion regions was selected using MPI/OpenMP parallelism, the speed of 6,144 cores reach 1,843,200 burnable regions per minute. Results from these studies demonstrate that RMC, using the hybrid parallel model described in this paper, is able to handle pin-wise full core burnup calculations with millions of depletion regions and scales well through thousands of processors. (authors)

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗

CTF Parallel Performance Improvements

CTF is a modernized version of COBRA-TF (Coolant Boiling in Rod Arrays-Two Fluid), which is a thermal hydraulic (T/H) sub-channel code based on a two-phase, two-fluid model that formulates the conservation equations of mass, energy, and momentum for three fields of vapor, continuous liquid, and entrained liquid droplets. Under the Consortium for Advanced Simulation of Light Water Reactors (CASL) program sponsored by the U.S. Department of Energy, CTF is being further developed and improved for modeling and simulation of light water reactors. The code improvements for predicting T/H responses of pressurized water reactor (PWR) cores are focused on its applications to specific challenge problems deemed significant to the industry, such as departure from nucleate boiling (DNB) and crud-induced power shift. The code is also being used for T/H feedback in normal operating condition depletions in PWR and boiling water reactor designs in the reactor core simulator, Virtual Environment for Reactor Applications Core Simulator (VERA-CS), being developed by CASL. CTF modifications under the CASL program include software optimization, new closure models, and parallelization for modeling full reactor core T/H responses. The purpose of this work was to improve the parallel performance of CTF. CTF was originally parallelized via a domain-decomposition strategy using the Message Passing Interface (MPI), in which the basis of a solution domain was a physical assembly in the model. In other words, a user who wanted to perform a parallel solution with CTF had to provide the code one processor per assembly in the model (e.g., 56 cores for a 56 assembly quarter-core model, or 193 cores for a 193 assembly full-core model). This design was rigid, demanding a very specific number of processors for a model; furthermore, when CTF was run in VERA-CS-which typically uses thousands of processors- many processors sat idle during the CTF solve because CTF was unable to use any more processors than there are physical assemblies. This was not necessarily a problem if the CTF execution speed was very fast, but the average time spent in a CTF solve for prototypical PWR operating conditions was significant (i.e., over 20 % of the simulation wall time was being spent in CTF for a standard core depletion simulation). Larger, more computationally intensive simulations performed in support of the CASL DNB challenge problem demonstrated a need for improvements in this original parallelization approach. The specific goal of this work was to allow CTF to use more processors than there are physical assemblies in the model, which would not only speed up CTF and improve the overall efficiency of VERA-CS but also give users more-flexible options for executing their simulations. (authors)

21 SPECIFIC NUCLEAR REACTORS AND ASSOCIATED PLANTS↗

Multi-fault Tolerance for Cartesian Data Distributions

Faults are expected to play an increasingly important role in how algorithms and applications are designed to run on future extreme-scale sys- tems. Algorithm-based fault tolerance (ABFT) is a promising approach that involves modications to the algorithm to recover from faults with lower over- heads than replicated storage and a signicant reduction in lost work compared to checkpoint-restart techniques. Fault-tolerant linear algebra (FTLA) algo- rithms employ additional processors that store parities along the dimensions of a matrix to tolerate multiple, simultaneous faults. Existing approaches as- sume regular data distributions (blocked or block-cyclic) with the failures of each data block being independent. To match the characteristics of failures on parallel computers, we extend these approaches to mapping parity blocks in several important ways. First, we handle parity computation for generalized Cartesian data distributions with each processor holding arbitrary subsets of blocks in a Cartesian-distributed array. Second, techniques to handle corre- lated failures, i.e., multiple processors that can be expected to fail together, are presented. Third, we handle the colocation of parity blocks with the data blocks and do not require them to be on additional processors. Several al- ternative approaches, based on graph matching, are presented that attempt to balance the memory overhead on processors while guaranteeing the same fault tolerance properties as existing approaches that assume independent fail- ures on regular blocked data distributions. The evaluation of these algorithms demonstrates that the additional desirable properties are provided by the pro- posed approach with minimal overhead.

Ali, Nawab↗

Measuring the capabilities of quantum computers

Quantum computers can now run interesting programs, but each processor’s capability—the set of programs that it can run successfully—is limited by hardware errors. These errors can be complicated, making it difficult to accurately predict a processor’s capability. Benchmarks can be used to measure capability directly, but current benchmarks have limited flexibility and scale poorly to many-qubit processors. We show how to construct scalable, efficiently verifiable benchmarks based on any program by using a technique that we call circuit mirroring. With it, we construct two flexible, scalable volumetric benchmarks based on randomized and periodically ordered programs. We use these benchmarks to map out the capabilities of twelve publicly available processors, and to measure the impact of program structure on each one. We find that standard error metrics are poor predictors of whether a program will run successfully on today’s hardware, and that current processors vary widely in their sensitivity to program structure.

97 MATHEMATICS AND COMPUTING↗

Fast and Accurate Simulation of the Cray XMT Multithreaded Supercomputer

Irregular applications, such as data mining and analysis or graph-based computations, show unpredictable memory/network access patterns and control structures. Highly multithreaded architectures with large processor counts, like the Cray MTA-1, MTA-2 and XMT, appear to address their requirements better than commodity clusters. However, the research on highly multithreaded systems is currently limited by the lack of adequate architectural simulation infrastructures due to issues such as size of the machines, memory footprint, simulation speed, accuracy and customization. At the same time, Shared-memory MultiProcessors (SMPs) with multi-core processors have become an attractive platform to simulate large scale machines. In this paper, we introduce a cycle-level simulator of the highly multithreaded Cray XMT supercomputer. The simulator runs unmodified XMT applications. We discuss how we tackled the challenges posed by its development, detailing the techniques introduced to make the simulation as fast as possible while maintaining a high accuracy. By mapping XMT processors (ThreadStorm with 128 hardware threads) to host computing cores, the simulation speed remains constant as the number of simulated processors increases, up to the number of available host cores. The simulator supports zero-overhead switching among different accuracy levels at run-time and includes a network model that takes into account contention. On a modern 48-core SMP host, our infrastructure simulates a large set of irregular applications 500 to 2000 times slower than real time when compared to a 128-processor XMT, while remaining within 10\% of accuracy. Emulation is only from 25 to 200 times slower than real time.

Cray XMT↗

Accurate, Fast and Scalable Kernel Ridge Regression on Parallel and Distributed Systems

Kernel Ridge Regression (KRR) is a fundamental method in machine learning. Given an n-by-d data matrix as input, a traditional implementation requires Θ(n2) memory to form an n-by-n kernel matrix and Θ(n3) flops to compute the final model. These time and storage costs prohibit KRR from scaling up to large datasets. For example, even on a relatively small dataset (a 520k-by-90 input requiring 357 MB), KRR requires 2 TB memory just to store the kernel matrix. Additionally, the reason is that n usually is much larger than d for real-world applications. On the other hand, weak scaling becomes a problem: if we keep d and n/p fixed as p grows (p is # machines), the memory needed grows as Θ(p) per processor and the flops as Θ(p2) per processor. In the perfect weak scaling situation, both the memory needed and the flops grow as Θ(1) per processor (i.e. memory and flops are constant). The traditional Distributed KRR implementation (DKRR) only achieved 0.32% weak scaling efficiency from 96 to 1536 processors. In this work, we propose two new methods to address these problems: the Balanced KRR (BKRR) and K-means KRR (KKRR). These methods consider alternative ways to partition the input dataset into p different parts, generating p different models, and then selecting the best model among them. Compared to a conventional implementation, KKRR2 (optimized version of KKRR) improves the weak scaling efficiency from 0.32% to 38% and achieves a 591x speedup for getting the same accuracy by using the same data and the same hardware (1536 processors). BKRR2 (optimized version of BKRR) achieves a higher accuracy than the current fastest method using less training time for a variety of datasets. For the applications requiring only approximate solutions, BKRR2 improves the weak scaling efficiency to 92% and achieves 3505x speedup (theoretical speedup: 4096x).

97 MATHEMATICS AND COMPUTING↗

Accelerating Scientific Computing in the Post-Moore’s Era

Novel uses of graphical processing units for accelerated computation revolutionized the field of high-performance scientific computing by providing specialized workflows tailored to algorithmic requirements. As the era of Moore’s law draws to a close, many new non–von Neumann processors are emerging as potential computational accelerators, including those based on the principles of neuromorphic computing, tensor algebra, and quantum information. While development of these new processors is continuing to mature, the potential impact on accelerated computing is anticipated to be profound. We discuss how different processing models can advance computing in key scientific paradigms: machine learning and constraint satisfaction. Significantly, each of these new processor types utilizes a fundamentally different model of computation, and this raises questions about how to best use such processors in the design and implementation of applications. While many processors are being developed with a specific domain target, the ubiquity of spin-glass models and neural networks provides an avenue for multi-functional applications. Furthermore, this also hints at the infrastructure needed to integrate next-generation processing units into future high-performance computing systems.

97 MATHEMATICS AND COMPUTING↗