High-performance computing tools for agent-based general ecosystems models (HPC-ABGEM)
PI: Christoph Zollikofer (University of Zurich)
Co-PIs: George Lake
January 1, 2014 - December 31, 2016
High-Performance Computing (HPC) is becoming increasingly important in fields such as astrophysics, climate research, fluid dynamics, chemistry and materials sciences, as well as in biology. In the latter field, HPC is used to model phenomena at different levels of scale, ranging from molecules (e.g. protein and membrane structure) to entire organs (e.g. brain and heart function). Surprisingly, however, the HPC approach is only beginning to be applied to large-scale Systems Biology, i.e., at the level of organisms, communities, and entire ecosystems, and over evolutionary time scales. These biological systems are typically composed of discrete units, which have individual properties, and which interact with each other and with their environment in non-linear ways. The classical top-down (analytical) modeling approach, which consists in solving sets of partial differential equations describing the hypothesized properties of the system, is thus complemented with a bottom-up approach, where the systems properties are defined with a set of rules specifying the behavior of its basic units. In Cellular Automata (CA) the basic units are stationary, while in agent-based models (ABMs), the units are mobile, thus representing an additional level of complexity. ABMs have become increasingly popular in areas such as socioeconomics, cultural anthropology, human and animal behavior, urban and traffic research, but applications are restricted to relatively small numbers of agents and time steps.
Recently, it has been proposed that it is time to model all life on Earth, i.e., to model entire ecosystems on the basis of individual organisms and their interactions with one another and with their environment. The potential and significance of implementing a Generalized Ecosystems Model (GEM) are enormous, but so are the computational challenges. Large-scale simulations involving realistic agent numbers for entire ecosystems complex agent-agent and agent-environment interactions, an extended number of time steps, and a sizeable parameter space exploration and replicas count, require a software architecture specifically designed for efficient parallelization and scalability. Clearly, a novel HPC-ABGEM (High-Performance Computing Agent-Based General Ecosystems Modeling) approach is required to meet these challenges. In a previous HP2C project (Modeling humans under climate stress), we implemented a prototype HPC-ABGEM simulation framework, which permits individual-based simulation of the dispersal of human populations interacting with their changing environment. The basic question we were asking was: how do small-scale/short-term interactions of humans with each other and with their environment result in large-scale/long-term system properties such as population dynamics and evolutionary change over time? We applied the code to simulate two well-known scenarios of human evolution: the replacement of Neanderthals by modern humans toward the end of the Pleistocene, and the global dispersal of modern humans over the past 50000 years. While the prototype simulation framework has already proven its potential for the simulation of human dispersal and extinction scenarios over evolutionary time scales, it has also revealed various limitations, which can only be overcome by a dedicated co-design approach. Current challenges are as follows: (1) ABMs typically involve multiscale systems made of heterogeneous, non-linearly interacting components. The resulting scalability issues represent challenges to agent design, memory allocation and data communication paradigms. (2) Agents are mobile units, which raises questions of spatial tiling, load balancing, and communication overhead. (3) Even HPCABGEMs are not intended to model every single organism of an ecosystem. Besides computational limitations, overparameterization tends to reduce the significance of in-silico modeling experiments. The challenge here is to design mixed models, which use a knowledge-based approach to combine bottom-up ABMs and CAs with top-down equation-based models.
Our long-term goal (10 years) is to establish HPC-ABGEM as a standard research tool for application domains such as ecology, evolutionary biology, biological anthropology, and socioeconomics. HPC-ABGEMs do not exist yet, but there is a clear need for them. Using this long-term perspective as a guide, the specific aim of this three-year PASC project is to co-design, implement, optimize and test a state-of-the art HPC-ABGEM framework. While we will use our existing code as a basis for further developments, the project goals can only be reached with a co-design approach, i.e., through hard-aware and soft-aware implementation of the simulation framework. Expert input from areas where the HPC approach is already well established will be instrumental; this input is provided via the astrophysics and biology networks. On the other hand, we expect to contribute to these networks with innovative concepts and implementations of highly parallelized ABMs. At the end of this project, a stable version of the HPC-ABGEM code should be ready to be distributed to the wider scientific community.