5
Computational Sciences
The Panel on Information Science at the Army Research Laboratory (ARL) conducted its review of ARL’s advanced computing architectures, computing sciences, data-intensive sciences, and predictive simulation sciences at Aberdeen, Maryland, on June 15-16, 2015. This chapter evaluates that work, recognizing that it represents only a portion of ARL’s Computational Sciences Campaign.
ADVANCED COMPUTING ARCHITECTURES
Alongside a focused research effort in tactical high-performance computing (HPC), researchers have moved to explore new and interesting computer architectures, including neuro-synaptic computing, the epiphany many-core chip, and quantum networking. The research staff in these areas of work exhibited high quality research skills and is well organized and positioned to succeed.
Within computational sciences, big data/analytics and HPC are treated separately. Given that important applications will increasingly combine high-performance analytics and high-performance computing, it is important to assess the impact of integrated computing on architecture and operating environments. Architectures will move to datacentric designs where compute is moved to data throughout the systems environment. This will be true in the data center and at the tactical edge. Analytics on very large dynamically changing graphs will be critical to gaining insights from big data. While the computation in scientific computing can be statically scheduled, computing on graphs is data-dependent. The operating environment therefore needs to change to schedule dynamic computations; this capability is also required for resilience in the battlefield.
Accomplishments
Work in tactical HPC has identified key aspects of the solution space and is exploring important issues related to establishing the utility of those aspects. A particularly interesting observation by ARL researchers is that the tactical HPC space is potentially richer than the commercial space, owing to differences such as communication reliability and potential deployment of field resources. The commercial space currently consists of handheld devices and cloud-computing resources deployed in data centers. The tactical HPC space includes handheld devices, field-deployed tactical (potentially mobile) resources, strategic resources close to the battlefield environment, and cloud-computing resources located in military data centers. ARL has developed models that will guide offloading decisions in these complex computing environments. An interesting question for ARL to explore is the commercial applicability of the intermediate resources to issues such as smart highways.
An example of tactical HPC is the project in multiobjective geometric optimization, a tool to safely guide soldiers to strategic viewing points in the presence of adversaries at known locations. The computational effort (ray tracing) to determine, in real time, a safe path in a 3D modeled megacity, avoiding detection by one or more known enemies, is significant and beyond the capabilities of current and emerging handheld devices. Input from the handheld device (tablet) to the remote HPC system and receipt of output back to the tablet do not place excessive demands on the energy-limited tablet. Researchers have explored a variety of solutions on a set of HPC platforms, with varying performance. This is a high-priority Army application, but security and resilience have yet to be considered.
The work on computation ferrying is exploring an important aspect of realizing ARL’s vision of tactical HPC. It considers the tasks that could be computed on handheld devices or on mobile tactical HPC resources. The project has developed a simulation environment in which to study the behavioral properties of the system to determine the specific scenario parameters for when it is necessary and reasonable to use offload tasks from the handheld devices. Initial results have demonstrated that offloading proves most useful for intermediate task sizes, which achieve a balance between communication cost and improved computation speed. Overall, this direction will motivate an end-to-end solution for battlefield computing.
In addition to establishing modeling and simulation capabilities to guide offload decisions, ARL has explored critical issues related to programmability and performance of edge (handheld) devices. The work with open computing language (OpenCL) programs demonstrates that devices can achieve significant computational efficiency, particularly in terms of power and performance. This work can serve as the basis for parameterizing the models and simulations that will guide offload decisions. A possible direction for future work on handheld devices is to develop an understanding of the trade-off between performance per watt (i.e., power efficiency) and the amount of power available in handheld devices (i.e., battery lifetime of those devices).
The project in dynamic binary translation (DBT), which demonstrated the fastest cycle-accurate emulator design in the 2014 MEMOCODE1 contest, falls in the areas of power-performance and heterogeneous computing. The research group developed a DBT to allow fast cross-architecture execution of binary codes. This would, for example, allow legacy binary code developed on a now obsolete architecture, e.g., i8080, for which the source code is not available, to be run efficiently on another current architecture, e.g., a low-power ARM.2 DBT is a well-mined area with results in the literature dating
___________________
1 MEMOCODE is the Association for Computing Machinery (ACM)-Institute of Electrical and Electronics Engineers (IEEE) International Conference on Formal Methods and Models for System Design.
2 ARM is a family of instruction set architectures for computer processors developed by British company ARM Holdings, based on a reduced instruction set computing (RISC) architecture.
back to the 1990s. What sets this work apart is that this DBT outperformed all other MEMOCODE entries by a factor of seven, with close to 1.1 cycle efficiency.
Challenges and Opportunities
In advanced computing architectures, the Computational Sciences Campaign specifically identified the goal of leadership in areas that include heterogeneous computing; many-core integrated architectures; tactical HPC; power, performance, and portability; quantum computing; neurosynaptic computing; and software-defined networking (SDN). Leadership is a posture in which ARL maintains considerable in-house expertise, has a substantive infrastructure, and devotes significant investment based on unique Army needs. If one adds to this objective a leadership role in the research community, then the above list is extremely ambitious. It would be sensible for ARL to define a more limited list of areas in which it will achieve this leadership.
ARL is developing in-house expertise with potential to establish a worldwide leadership role in neurosynaptic computing and has plans to establish substantive infrastructure by investment in the area. Additionally, it is investing in quantum networking as a specific focus within the topic of quantum computing.
There is significant research effort worldwide in heterogeneous computing, many-core integrated architectures, and power, performance, and portability. Most of this work is taking place within data centers, with some researchers exploring deployment in the field. ARL has a clear opportunity for broader community leadership and significant impact on the Army’s capability if research in these areas is focused on tactical HPC. As part of this work, ARL could investigate advanced techniques for security and resilience in the battlefield infrastructure. The battlefield environment implies much higher costs for usability and reliability than does a commercial environment. The investigation here could look at the rescheduling of work dynamically as parts of the infrastructure go offline and come back online.
The majority of the posters presented as part of advanced computing architectures can be brought under this umbrella. While the posters mostly represented good work, many were missing the driving application or scenario, and they tended to be incomplete. The two posters related to SDN (software-defined networking) were missing a critical piece—the dynamic reconfiguration of the network has to be driven by the requirements of the workload, which need to be specified in declarative language or detected automatically. The ARL researchers are to be commended for their participation in the GENI (global environment for network innovations) community and for the quality of the infrastructure that they have built.
The two key challenges for tactical HPC are (1) the selection of high-priority Army applications and scenarios to drive the research and (2) the inclusion of security and resilience of the infrastructure as an explicit research goal. The latter challenge may necessitate new hiring and/or the forging of external partnerships.
Broader community leadership in quantum networking is also possible, but a more in-depth review of this thrust needs to be considered. The research is being done in the context of a larger Center for Distributed Quantum Information that has the specific objective to study the essential elements for implementing and exploiting a resilient network of quantum devices. The level of resources allocated needs to be reassessed if the goal is to build leadership in this area.
ARL researchers are pursuing interesting and important research in the area of advanced computing architectures. However, the work frequently is not appearing in tier one conferences and publications—e.g., International Symposium on Computer Architecture; IEEE’s Micro’s Programming Language Design and Implementation (PLDI), Architectural Support for Programming Languages and
Operating Systems (ASPLOS), and SC (formerly Supercomputing); and the International Conference for High Performance Computing, Networking, Storage and Analysis, which are the most important publication venues for research in the area. The key issue for achieving such publication successes for ARL is learning how to package research results for those venues. ARL researchers can only learn this skill by attending those conferences and participating in their organization, such as program committees.
COMPUTING SCIENCES
Computing sciences aims to develop the understanding, tools, techniques, and methodologies to fully exploit emerging computing architectures. This is accomplished through the realization of efficient task parallel algorithms and the use of advanced memory hierarchies. These efforts are expected to greatly reduce the time required to restate algorithms in parallel form and to correct implementation faults and bugs. The research areas of focus can be broadly categorized as programming languages, computational environments, and software integration. More specifically, research in programming languages seeks to improve performance, portability, and productivity of methods for Army-specific applications. Software integration aims to reduce the cost of entry for developers of scientific software and for scientists, engineers, and other users, allowing them to effectively use evolving computing environments. The evolution of the computing infrastructure is creating new challenges, ranging from energy-aware software development to software for massively parallel and distributed systems. The research in software integration is focused on addressing these challenges and developing an approach for software design that is predicated on the principle of systematic software reuse.
Accomplishments
The ARL computing sciences group has established a strategic focus in quantum computing; parallel processing environments for large, heterogeneous parallel systems; and tools to simplify application development for HPC environments. Establishing a coherent and sustaining strategy in these areas was a major accomplishment. The group has also grown its work in exploratory research while still providing support for issues of immediate importance to the Army’s mission. The group has established an approach to keeping its computing infrastructure up to date with the latest hardware and software systems and tools, effectively upgrading the HPC system infrastructure every 2 years (staggered replacement of two HPC systems, each on a 4-year replacement cycle). While this turnover in equipment can be a challenge for application developers, the group is working to provide tools to simplify application development and help the developer optimize performance on the more advanced systems.
Research in quantum computing and software environment optimization is especially noteworthy and recognized as being leadership work that advances the basic science in important areas of computing technology. The focus within quantum computing is hardware abstraction at the function level (referring to a network of devices rather than an individual device) versus the Qbit, or device, level. This work has the potential to advance the knowledge and technology for future quantum computing systems.
Similarly, the focus on software environment optimization for hierarchical many-core and heterogeneous processing environments is on tools and algorithms to support code development and optimization of emerging HPC system structures. This work will be important for the efficient and effective use of HPC systems under development for ARL applications.
Two projects stood out as incorporating outside-the-box concepts and resulting in high-quality basic research that integrated theory, computation, and experimentation. These projects were focused on the development of threaded message-passing interface (MPI) for reduced instruction set computing (RISC)-
array multicore processors and on HPC scaled quantum hardware description language (QHDL)-based modeling of entanglement dynamics in Jaynes-Cummings circuits in open system evolution. The first of these provides ARL a solid foundation to pursue innovative solutions to the challenging problem of power-efficient parallel programming. The second is one of the few efforts in quantum networking, and the development of hardware abstraction language will have strong value in future systems. Another project, dealing with an auto-tuning benchmark for HPC accelerators, represents valuable tactical research versus long-range basic science. This research supports initial performance optimization of existing applications and benchmark codes on emerging HPC systems. Auto-tuning in this work is a higher level tuning (as contrasted with a widely practiced industry approach of using libraries).
Work related to communication-avoiding approaches for Lanczoz eigensolvers demonstrated a good understanding of the interaction of mathematics and HPC. The approach is in the process of being submitted to Sandia National Laboratories for inclusion in production codes (Anasazi) to realize the benefits of scaling; the potential for such deployment is a sign of high quality and immediate scientific relevance and impact. A project dealing with hierarchical multiscale computational framework was technically strong but not considered as representing basic research. This work could be transitioned to a production code.
Challenges and Opportunities
The major challenges facing the computing sciences group include scaling applications across large parallel heterogeneous systems, developing tools and frameworks that work on multiple parallel system architectures, and improving software performance and developer productivity in the face of all evolving system architectures and capabilities. System designers have to create more exotic system architectures to increase system capability and performance, since processor clock speed has been flat for many years. Also, new and emerging applications target the processing of very large data sets or cover a large problem space (e.g., time and structure size), putting pressure on system designers to consider system architectures optimized for datacentric computing (large data/computation ratio). The complexity of system architectures and the need to handle large data sets have significantly increased the challenge of giving application developers effective and efficient software environments, frameworks, and languages. However, ARL’s computing sciences group has the opportunity to create the tools needed to optimize performance and scale for mission-critical applications and user environments. Increasing the capability of the Army is dependent on the computing sciences group’s development of technologies to continue scaling of capacity and performance of the Army’s computing infrastructure and emerging applications.
Research related to domain-specific languages is an appropriate focus given the promise of improvements in productivity. Work is ongoing to create a new language (e.g., Liszt) for domain-specific parallel programming. The analysis of results did not include comparison with other graphics processing unit (GPU) languages, and the metric used to evaluate productivity (e.g., lines of code) was not sufficient. It was not clear if this work holds promise of transitioning to an ARL customer or to any outside agencies, given the lack of control of GPU architectures. It was also unclear whether work on datacentric, extreme-scale computing provides a viable, sustainable, or unique capability over existing task-oriented parallelism frameworks. Related work on adaptive fast multipole methods using task-oriented parallelism suffered from a lack of supporting performance data and insufficient analysis and articulation of scaling properties. Work related to the implementation of a Bayesian quantum game builds an experimental, computational, and theoretical framework for using quantum statistics to model aspects of cognition. While the approach is conceptually promising, a more detailed presentation and references to related work would be beneficial.
DATA-INTENSIVE SCIENCES
The goal of the data-intensive sciences research portfolio is to understand and exploit the fundamental aspects of large-scale, multidimensional data analytics. There are currently three areas of focus in data-intensive sciences research: the science of large data, computational mathematics for data analytics, and real-time data access and analytics. The first two of these research areas were part of the current review.
Accomplishments
Current accomplishments in this research portfolio include new model order reduction methods for partial differential equations (PDEs), cognitively steered exploration of solutions to PDEs at different resolution, efficient summarization and visualization of high-dimensional data sets, and concise characterization of spall damage in materials. ARL has a goal of leading research in the areas of explosive mechanisms and high strain rate and fracture. A new research effort has been prepared in computation of game strategies exploiting symbolic representation and in using a new neuromorphic computational architecture.
It is encouraging to see the focus in these important areas. The initial set of projects appears to be promising and is beginning to show good progress. Specifically, the work on developing a high-performance, sparse, nonnegative, least square solver advances the state of the art and leverages ARL’s expertise in numerical analysis and HPC. Similarly, the work on neuromorphic computing represents a fresh and original approach.
The work related to model order reduction methods for large-scale simulation data is part of a larger collaboration with the Stanford University group on model order reduction. ARL is pioneering the parallelization of hyperreduction methods to bring greater computational efficiency in the simulation of PDE codes. The approach being implemented is novel and important. The researchers are well informed on the state-of-the-art approaches in this field, and appropriate equipment and numerical models have been selected. The availability of world-class HPC resources is a strong driver, and the collaboration with the Stanford University group is impressive. The work contains the appropriate mix of theory and computation. Additional work in deriving error estimates and bounds for the hyperreduction results would strengthen the impact of this work.
The new thrust in programming approaches for neuromorphic cognitive computing represents a novel and intriguing research program. The principal investigator has the appropriate background for this work and has demonstrated a good understanding of the fundamental research questions. The use of hand-coded basis functions is appropriate for the initial stages of investigation, but other alternatives could be considered that have a basis in coding theory—for example, the use of low-density parity check codes. This research thrust is at a preliminary stage, and it is too early to decide whether an accelerated resource track would result in a high-impact transfer.
A research effort related to large-scale network experimental data reduction has led to acceleration of processing time for measured network packet traces. The approach uses a standard relational database with a combination of custom query processing that includes map-reduce and data summarization and compression. Providing several new visualizations of network operation augments the approach. The work is of high value and is adequately funded. One possible extension of the work would be to advance the state of the art for visualization through the use of nonlinear manifold learning techniques and unsupervised explanation of system regimes to reveal anomalous phenomena.
The research portfolio includes a focus on large multiscale material modeling with a special emphasis on identification of damage modes. Molecular dynamics simulations have been used in conjunction with a sectioning data reduction approach to develop a new and elegantly simple hydrostatic criterion for spallation of crystalline materials. The new criterion was then used to visualize the stress state from the data set obtained for an impact simulation. The research provides a new characterization of spallation. Furthermore, the approach, using molecular scale simulation coupled with novel data characterization, is broadly applicable in related problems.
The work of the visual simulation laboratory is focused on the use of a visualization-based framework to allow users to steer a multiresolution PDE simulation. The principal contribution of this work is in the computational steering of a parallel algorithm, but it was not clear whether the innovation was in the parallelization approach or in the visual framework in which the simulation was performed.
Challenges and Opportunities
Currently there is no research roadmap that identifies specified projects and goals linking research efforts to the stated goals—the science of large data, computational mathematics for data analytics, and real-time data access and analytics. The portfolio of research is primarily focused in computational simulations on HPC platforms. The portfolio could be broadened considerably given ARL’s varied and demanding real-world data intensive applications that go well beyond computational simulations. Some examples are big data analytics in the context of social networks, geographic information systems, and situation awareness (e.g., in the context of cybersecurity). There is an opportunity to take a leadership role in developing scalable numerical optimization methods and multiscale and linear algebra methods for state-of-the-art machine learning that are suitable across a wide range of architectures, including large HPC systems, tactical HPC, and mobile systems. Existing expertise and strengths in the areas of numerical analysis, scientific computing, and parallel computing need to be leveraged in this context.
The data sets of interest for Army applications are characterized by their complexity and high dimensionality, and computing constraints often dictate computations that are approximate or statistical in nature. The data-intensive sciences research could be strengthened by incorporating more rigorous problem formulations that take such constraints into account and that lead to results with theoretically grounded error bounds. The research would contribute to ARL focus on verification and validation in computational science and would also enhance visibility through improved quality of publications. Given that many applications will combine big data, HPC, and high-performance analytics, there is a tremendous opportunity for datacentric computing to impact the design of HPC systems and their operating environments.
ARL is a Department of Defense (DOD)-wide supplier of supercomputing capacity and networking. The state of the art has today shifted to an emphasis on data-rich computations. In many sciences, large, widely accessible databases—for example, protein databases and large language corpora—have played a fundamental role as engines for progress. Military-focused computations can benefit today from such an approach. For example, huge data sets comprising network data, textual interactions, and even audio logs are generated and archived by large training exercises. The collection, curation, and dissemination of such large military-focused data sets could be vital to the ARL research mission as well as future engineering and technology efforts throughout DOD. This is a natural evolution for ARL. Serving these data could be over the same communication networks that ARL currently maintains. The data could be housed in the same supercomputer centers that ARL operates for DOD. Ultimately, such a service would be far more valuable than just providing raw compute cycles.
It is important for ARL researchers to anticipate new technologies and how they could impact its research agenda. One example of a technological trend to monitor is the emerging disruptive revolution in memory technology. Akin to the effect that cheap high-volume, solid-state drives (SSDs) have had on the supercomputing landscape, a larger disruption is expected with the high-speed, large-volume nonvolatile memories that will be available in the near future. How this will affect architectures and especially algorithms and software design is unknown today. ARL could take a leadership role in understanding the impact in general and especially with regard to Army applications.
Similarly, ARL research could also consider a focus on the missing elements of a computational architecture that could provide battlefield information dominance architecture (IDA). ARL is contributing strongly to this goal in the front end, via advanced sensing and networking. For the middle layers of an IDA, where data are reduced, fused, correlated, analyzed, and reduced to actionable information, there is evidence that ARL is building leading capabilities and advancing the state of the art. However, it would be desirable for ARL to devote more attention to looking at the back end of the IDA, where information is delivered to the individual warfighter. An information chain is limited by its weakest link. Beyond human factors research, research into the optimal adaptation of new and emergent technologies required for architecting the back end is required. As an example, low-latency, high-fidelity virtual reality (VR) devices and rendering engines made possible by advances in computing and organic light-emitting diode (OLED) displays will reset fundamental user interface (UI) assumptions in a far more profound way than did two-dimensional (2D) graphical user interfaces (GUIs), mobile or touch. The implications of these devices for the back end of the IDA are equally profound. An opportunity exists for ARL to take a leadership role in understanding how advanced displays, visualization techniques, and algorithms can be used, beyond the current applications to training, to reach an integral part of the future battlefield.
PREDICTIVE SIMULATION SCIENCES
The research program in predictive simulation sciences reflects an appropriately broad understanding of the underlying science and of comparable R&D activities at other institutions and agencies. Many of the research projects include collaborations with academia, federal laboratories, and industrial partners, and a few already have transition plans in place to deploy the results of the research in appropriate national security settings. Most of the publications cited are of recent vintage, and many have academic or industrial collaborators. A review of the work shows a few gaps in broad understanding, largely explained by the nature of the work, which requires expertise in both computational sciences and in the application domain. This underscores the importance of collaborations in the work.
Numerical models are generally relevant to the work proposed. Many of the projects are currently compute-bound and are now exploring ways to more effectively utilize additional computational capabilities available via GPUs or other exascale technologies. Close collaboration with leading universities and federal laboratories actively working on exascale computational technologies needs to be encouraged in ARL’s Computational Sciences efforts to ensure that state-of-the-art developments in these fields can be transferred to ARL’s Computational Sciences R&D projects in a timely manner.
The research qualifications of the ARL staff are relevant to the technical challenges inherent in the R&D work. There are some indications that the cultural changes necessary to effect a transition to a campaign-based strategy for R&D initiatives are still in progress, but no substantial expertise gaps were observed in the predictive simulation sciences presentations and poster discussions.
The R&D efforts presented incorporated appropriate levels of theory, computation, and experiment. Experiment could use existing validation data to ensure that simulations are predictive, but some of the research utilized real-world laboratory results to steer or refine computational science predictions.
For example, the meso- and microscale weather project utilized airborne weather sensors to improve microscale models in near real time. Additional effort in integrating theory, computation, and experiment via formal validation and verification (V&V) techniques would be a welcome addition to many of these projects.
Some exceptionally promising projects found in the Computational Sciences Campaign R&D portfolio are listed separately below. Most projects showed evolutionary progress, but there are projects that may provide revolutionary benefits to the ARL mission.
Future outside-the-box, close collaborations between predictive simulation science R&D projects and information sciences efforts have the potential to create novel hybrid technologies from the fusion of these two technology arenas. Large-scale computational simulation and large-scale data analytics seem an especially appropriate fusion for ARL R&D investigations.
Accomplishments
Multiscale material modeling is a potentially game-changing computational technology for predictive simulation in the mechanical sciences. The work on multiscale simulations based on scale-bridging combines several important threads of cutting-edge research into a useful software product that has the potential to provide higher fidelity deterministic and nondeterministic forward simulations in fluid and solid mechanics. These multiscale simulations are essential for assessing vulnerability, lethality, and effectiveness of weapons and protection systems, and the current effort demonstrates the project’s utility in theory and also in practical application to software commonly used by DOD (e.g., Lawrence Livermore National Laboratory’s ALE3D production software tool,3 used for high-explosive weapons and target simulations). In addition to this research being of high value for predictive forward analysis, its multiscale sampling subsystem can lead to vastly improved performance for inverse analyses and quantification of margins of uncertainty (QMU) estimations. The latter, in particular, are often computationally expensive for application to large-scale problems. The principal technical challenges for this work lie in the availability of HPC resources, because the smaller scale simulation components can consume prodigious numbers of computational cycles. The research group has worked to overcome this challenge by deploying an optimal sampling scheme to reduce computational burdens on the smaller scale simulation components and by utilizing software broker components to aid in the load-balancing required for optimal use of HPC resources.
An important focus of the ARL research related to developing predictive capabilities for use on low-power computer platforms representative of what would be available to Army personnel in the field. Research on synchronous time-driven simulation for HPC accelerators focused on the simulation of large ensembles of random events for tactical cybersecurity applications. The research approach was to simulate the continuous real-world setting by utilizing a finite-state approximation that would converge to the continuous case as time step size decreased. Synchronization of events is problematic, with longer time intervals and smaller time step size, causing computational costs to increase significantly. The focus is to seek statistical convergence for the larger time steps. Validation of results on larger and more realistic data sets would be warranted.
In the area of materials modeling, research on scalable algorithms for simulating dislocations in microstructured crystals focuses on fusion of discrete dislocation theory coupled with continuum finite-element approximation, toward the capture of inelastic effects in polycrystalline and heterogeneous mate-
___________________
3 ALE3D is a 2D and 3D multiphysics numerical simulation software tool using arbitrary Lagrangian-Eulerian (ALE) techniques.
rial structures. The research is promising and has broad applicability for material and structural failure simulations. Each element of this algorithmic fusion has strengths and limitations, but the coupling of these two techniques promises to remedy the weaknesses of each while preserving their considerable strengths. One of the most difficult technical challenges for this work is the effective load balancing of HPC resources. In computational physics, localization phenomena are not handled effectively through static load-balancing techniques, and the research could benefit from consideration of new developments in adaptive parallel load-balancing techniques. Because a similar challenge arises when transitioning from current computing architectures to emerging pervasively threaded hardware such as GPUs, due attention to research and development efforts in academia and other federal research institutions efforts is warranted.
The research on computational drug discovery focuses on developing efficient techniques for discovering which readily available molecules may be suitable for synthesizing drugs for use against pathogens that an adversary might deploy as biological weapons. This effort leverages a variety of university collaborators and is helped by access to ARL HPC resources such as the Excalibur supercomputer. The broad impact of the success of this research is significant for national security and in public health settings. Considerations relevant to technical limitations of this work include whether currently available libraries contain potential drug candidates, whether the identified compounds are optimal in effectiveness, and whether limiting the scope of these searches to small compounds is the best path for drug discovery.
Challenges and Opportunities
Predictive simulation sciences are increasingly important in DOD R&D venues. A broader understanding of the strengths and weaknesses of modeling and simulation techniques and of their impact on Army R&D practices is essential to help steer ARL research activities toward full realization of future simulation trends.
A challenge that cuts across the predictive simulation sciences portfolio is the lack of cutting-edge R&D efforts in validation, verification, and uncertainty quantification. V&V are indeed essential for all simulation applications, and failure to attend to these basic measures puts all of ARL’s computational effort at risk of being of little value in the decision-making process. It is therefore important to raise V&V and related research in QMU as an important intellectual thrust in the ARL campaigns.
Many of the presentations and posters referred to an interest in utilizing emerging hybrid (i.e., thread-scalable with distributed memory, commonly termed MPI+X) computational architectures. The researchers at ARL who work in these intellectual venues can learn much by leveraging the considerable ongoing R&D efforts in advanced programming models for exascale computing that are occurring in leading universities and federal laboratories. In addition to these promising HPC developments, many of the computing-at-the-tactical-edge projects could also benefit from closely following these ongoing R&D efforts. One promising side benefit of exascale HPC research is the development of lower-power petascale and terascale resources that would be appropriate for this class of tactical computing.
The use of computational fluid dynamics (CFD) tools for the aerodynamic analysis and prediction of flight behavior of Army ballistic weapons parallels similar work in academic, industry, and federal laboratory settings. The coupled aerostructural simulation tools provide useful preliminary analyses at low Mach numbers. A serious limitation of the ongoing work is the assumption of rigid-body response for the projectile. While such an assumption would greatly simplify the computation, its validity has to be questioned for slender projectiles. Furthermore, while the assumption of rigid-body dynamics would hold for conventional artillery projectiles, finned structures are often sufficiently flexible and would impugn the accuracy of the predictions. Additional effort needs to be directed at extending the
methodology for larger scale problems, higher resolution analyses, and optimizing the computational codes to take full advantage of emerging computational architectures.
OVERALL QUALITY OF THE WORK
There has been a significant improvement in quality of the research over the past 3 years, as indicated by the methodology pursued, the quality of the personnel assembled, and the infrastructure capabilities that have been developed over this period. The majority of the research presented (in both scheduled presentations and in the posters) was of comparable quality to that conducted in leading research universities or in federal and industrial laboratories. There were some projects where researchers clearly demonstrated a deeper knowledge of existing theory, of prior research in the field, and of state-of-the-art technology in relevant areas. Other projects provided a clear problem statement, including the articulation of a specific strategic direction for the research. It would be important to mentor researchers so that these best practices are embraced more broadly across all research groups. In each R&D effort, due emphasis was given to publication in relevant journals or proceedings, including (but not limited to) the open literature. Collaborations with university researchers via the ARL open campus initiative would provide opportunities for ARL staff to gain experience in the professional value of a continuous publication record.
Over the next decade, big data, complex analytics, and modeling and simulation will all come together in critical ARL applications and environments. This is evident from the materials provided by ARL in the area of computing sciences, data-intensive sciences, and predictive simulation sciences. Many of these applications will execute in a battlefield environment, where security and resilience will be of paramount importance.
To gain transformational improvements in power, performance, and resilience, the entire stack—hardware, software, and applications—needs to be codesigned. ARL has a tremendous opportunity to be a leader, especially in battlefield environments. To accomplish this, ARL would benefit from tasking an enhanced team of technical leaders to ensure that the subareas are increasingly engaged in real codesign.
The Computational Science Campaign is an inherently interdisciplinary enterprise, providing enabling technologies that cut across all of the ARL R&D campaigns. The capabilities of the Computational Sciences Campaign support virtually the entire spectrum of ARL R&D efforts. A broad swath of the projects could benefit from a better understanding of the theoretical and mathematical foundation that governs these problems. An increased contact and collaboration with researchers strong in theory—for example, from mathematics and computer science as well as the foundational science domains—would help achieve this. In particular, adding more computer scientists to the computational sciences research effort would encourage a more rigorous approach and would help advance the research agenda.
The projects reviewed comprised a mix of long-range basic science research and short-range research addressing tactical challenges and opportunities. The short-range research work is important, given the many challenges faced by the Army in its fields of engagement; ARL is developing innovative solutions to these high-priority tactical problems. It is also critically important that ARL continue to bring greater emphasis in its research portfolio to long-range strategic research in advancing fundamental knowledge in the underlying science. Strong relationships with key external entities have proven to be valuable across the breadth and depth of ARL research activities. This is an important positive development for ARL, and engagement with additional leading research institutions would further strengthen the awareness and capabilities of ARL researchers in important strategic areas.
