EAGER: Myriad: a new architecture for parallel multiscale simulation on CPU/GPU

科学知识的扩展正在不断提高人类对大脑和其他生物系统的理解。然而，越来越清楚的是，大脑的许多能力来自其许多区域之间丰富而复杂的相互作用。相反，大脑和身体的疾病往往是由细微的缺陷引起的，这些缺陷逐渐损害了这些相互作用的神经和生物系统的协调功能，直到最后，它们开始崩溃并表现出主要症状。由于这些相互作用的网络太过复杂，无法简单地凭直觉判断，因此科学家们构建了计算机模型，以便将许多不同的科学研究成果汇集在一起，并逐渐构建成对所研究的复杂系统的统一理解。这使得我们对复杂的大脑和身体有了更广泛的理解，如果没有这样的定量模型，这是不可能的。Myriad项目的目标是为生物建模提供一个软件平台，使利用现代并行处理计算机系统的能力变得更加容易。第二个目标是实现一个模拟器，使用这个平台来创建神经元和神经网络的详细模型。Myriad是一个基于共享内存架构的分区模拟器平台，设计用于基于NVIDIA GPU （CUDA）或并行CPU执行的计算速度。它的变革潜力来自于它自动并行化任何分区模型的能力，包括那些目前不能有效并行化的密集模拟交互，并且不需要最终用户编写特殊的、特定于平台的并行化代码。Myriad的共享内存设计避免了消息传递，并利用了一种彻底的颗粒设计方法，使分层定义的细胞模型扁平化，甚至可以通过状态变量分解单个等长间隔。具体来说，所有可以表示为由任意数量的任意机制连接的等距离、有状态节点（隔间）的模型都可以用高度并行性进行模拟，自动按线程数量缩放到可用线程数量，并以非常精细的粒度进行负载平衡，以最大限度地利用可用CPU或GPU内核。在编程上，终端用户模型在基于python的环境中定义，并通过代码生成技术转换为完全指定的C99代码（用于CPU或GPU），这些代码生成技术由自定义抽象语法树（AST）转换器增强，对于NVIDIA GPU，为CUDA提供了一个自定义对象规范，可以完全在卡上执行。Myriad的第一个应用将是模拟生物物理上真实的神经元和网络计算模型。然而，重要的是，Myriad的通用分区求解器原则上能够并行化任何可以表示为通过任意机制成对耦合的有状态节点的模型框架。因此，要将Myriad扩展到新的科学研究领域（例如，基因调控网络、流行病学模型、宿主-病毒相互作用、生态系统），只需要在基于python的较高软件层进行开发，可能需要相关领域的专家进行开发。最后，由于目前的并行化策略只能加速稀疏耦合模型，计算建模最近已经被吸引到能够通过这些方法解决的科学问题的子集。Myriad的发布可能会重新开启定量生物学问题的广度，这些问题可以通过并行模拟有效地解决。有关Myriad项目的最新信息可以在http://cplab.net/myriad.This上找到。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The expansion of scientific knowledge is continually improving human understanding of the brain and other biological systems. Increasingly, however, it is becoming clear that many of the capacities of the brain arise from rich and complex interactions among its many regions. Conversely, disorders of the brain and body often arise from subtle deficits that gradually impair the coordinated function of these interacting neural and biological systems until, at last, they begin to collapse and major symptoms are expressed. Because these networks of interactions are too complex to simply intuit, scientists construct computer models so that the work of many different scientific studies can be brought together and gradually constructed into a unified understanding of the intricate systems under study. This leads to broader understandings of our complex brains and bodies that would not be possible without such quantitative models. The goal of the Myriad project is to provide a software platform for biological modeling that makes it much easier to harness the power of modern parallel-processing computer systems. A second goal is to implement a simulator that uses this platform to create detailed models of neurons and neural networks. Myriad is a compartmental simulator platform based on a shared-memory architecture and designed for computational speed based on NVIDIA GPU (CUDA) or parallel CPU execution. Its transformative potential arises from its capacity to automatically parallelize any compartmental model, including those with dense analogue interactions that presently cannot be effectively parallelized, and without requiring the end user to write special, platform-specific parallelization code. Myriad's shared-memory design eschews message-passing, and utilizes a radically granular design approach that flattens hierarchically defined cellular models and can even break up individual isometric compartments by state variable. Specifically, all models that can be represented as isometric, stateful nodes (compartments) connected by any number of arbitrary mechanisms can be simulated with a high degree of parallelism, automatically thread-scaled to the number of available threads and load-balanced with very fine granularity to maximize the utilization of available CPU or GPU cores. Programmatically, end-user models are defined in a Python-based environment and converted into fully-specified C99 code (for CPU or GPU) via code generation techniques that are enhanced by a custom abstract syntax tree (AST) translator and, for NVIDIA GPUs, a custom object specification for CUDA enabling fully on-card execution. The first applications of Myriad will be to simulate biophysically realistic computational models of neurons and networks. Importantly, however, Myriad's generic compartmental solver will be able in principle to parallelize any model framework that can be represented as stateful nodes coupled pairwise by arbitrary mechanisms. Accordingly, to extend Myriad into new scientific areas of study (e.g., gene regulatory networks, epidemiological models, host-virus interactions, ecological systems) will require development only at the higher Python-based software layer, presumably by experts in the relevant field. Finally, as present parallelization strategies are only able to accelerate sparselycoupled models, computational modeling recently has been drawn towards that subset of scientific questions capable of being addressed by these methods. The release of Myriad may reopen the full breadth of quantitative biological questions that can be effectively addressed by parallel simulation. Updates on the Myriad project can be found at http://cplab.net/myriad.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.