By 
On June 25, the Intelligence Advanced Research Projects Activity (IARPA) posted a request for information for the System-Level Application Modeling (SLAM) program (IARPA-RFI-19-08). Responses are due by 4:00 p.m. Eastern on July 29.
The Intelligence Advanced Research Projects Activity (IARPA) is seeking information on research efforts in the area of future modeling and simulation (hereafter “ModSim”) research for large-scale computational and data-analytic applications. Thisrequest for information (RFI) isissued solely for information-gathering and planning purposes; this RFI does not constitute a formal solicitation for proposals. The following sections of this announcement contain details about the scope of technical efforts of interest, along with instructions for submitting responses.
Background & Scope
Modeling and simulation (ModSim) is essential to the co-design of new system architectures (hardware and software) and applications. It provides a cost-effective way to explore new designs, to examine the impact of excursions from a baseline design and to optimize the hardware and software before a final build.
The ModSim challenges posed by escalating system and application complexity are many and include: application execution, high-performance data movement, data management, computation scheduling, and representation of system characteristics. An application could involve a mix of streaming data and file system-based bulk data. This additional challenges of heterogeneous data sources make modeling the execution of an application an even more important, but complicated effort.
This RFI seeks to identify advances in the following research disciplines:
High-Fidelity ModSim Techniques.
New high-fidelity models of complex and heterogeneous architectures designed to execute large scale data-centric applications, under dynamic conditions of system performance and optimization require new modeling methods and tools. New methodologies must be flexible and low in computational demands. Simulation fidelity and accuracy must be high and provable or, at a minimum, consistent over a diversity of architectures and application workloads. Of particular interest is near real-time high fidelity simulation.
ModSim of Artificial Intelligence Systems and Applications and Machine Learning as a Method of ModSim.
Artificial Intelligence (AI), in general, and Machine Learning (ML), in particular, have emerged as essential application drivers in all forms of computing, including large-scale data analytics and numerically-intensive computations. The trend’simpact extends beyond the nature of architectures optimized for executing an ML workload. It also points toward applying AI/ML techniques as ModSim methodologies to support a range of systems (including but not limited to AI-centric systems).
Unified Modeling of Performance, Power, and Resilience.
As systems scale in size and technology moves toward highly-resilient and energy-efficient system designs, it is critically important to develop new integrated methods that capture performance, power/energy consumption, and resilience while considering the impacts of thermal effects and power constraints. For example, software-level resilience techniques, including algorithm-based fault tolerance, checkpointing, and architecture-level optimizations may be required to operate at low supply voltages. Such models would enable quantification of the trade-offs between performance, power/energy consumption, and reliability.
System-Level Modeling and Simulation.
As defined here, a “system-level model” would involve integrative ModSim technologies to predict performance and energy consumption of the whole system while the system is executing an application of interest. It is highly desirable to be able to design ahead and optimize before and after implementing the whole system using ModSim. The integration of ModSim methods must be over the subsystems and their interfaces and incorporate the ability to “dial-in” resolution and fidelity, depending on the simulation’s intended use. System-level modeling should cover diverse applications of interest with different complexity and workload characteristics. Such an integrated simulation could require mixing and matching methods based on different modeling approaches (e.g., analytical and statistical).
Modeling Irregularity in Applications.
Dynamic behavior stemming from both applications and system adaptivity will soon become an essential aspect in normal system operation. As execution models move toward more dynamic task-oriented models, predicting application performance in advance becomes even more challenging due to migratable work units and automatic load balancing mechanisms. Moreover, capabilities such as dynamic power steering, thermal throttling, and process variation result in performance heterogeneity across the system, rendering even carefully initially balanced workloads suboptimal. Capturing this variability is fundamental to advancing state of the art in ModSim techniques.
Dynamic Modeling.
ModSim methods that can quantitatively and accurately capture dynamic and adaptive applications and systems behavior are needed. Current leading-edge modeling techniques produce static models that target traditional scientific workloads and are therefore inadequate for addressing dynamic behavior. Existing proven methods must be extended to predict performance and energy consumption of applications whose behavior is input-dependent and dynamic throughout execution. Establishing a dynamic, adaptive modeling methodology would be a crucial enabling technology for efficient and productive future large-scale systems utilization.
Full information is available here.
Source: FedBizOpps
