high performance computing on graphics processing units: hgpu.org

Agent-Based Simulation (ABS) is a modelling approach where simulated entities i.e., agents, perform actions autonomously and interact with other agents based on a set of rules. ABSs have demonstrated their usefulness in various domains such as transportation, social science, or biology. Agent-based simulators commonly rely vastly on Central Processing Unit (CPU)-based sequential execution. As a result, they often suffer from long execution times, especially when the simulation model is complex and the scale is large. One approach to circumvent this problem is to employ heterogeneous hardware, e.g., systems equipped with both CPUs and Graphics Processing Units (GPUs). Although existing works have shown this to be a promising solution, simulators that support heterogeneous hardware are still not widely utilized. As this thesis demonstrates, blame can be laid on the difficulties to find a viable partition of simulation components to run on different hardware devices as well as steep programming requirements to enable ABSs on heterogeneous hardware. To ease the aforementioned issues, in this thesis, we present a full framework for automated generation of high-performance simulation code targeting heterogeneous hardware. The framework consists of two parts: OpenABLext and OptCL (short for Optimise performance targetting high-performance domain-specifiC Languages). OpenABLext significantly extends the existing OpenABL, a tool designed for generating high-performance ABS programs from sequential representations by parallelising the execution of so-called step functions which implement the agent-based models. We first extend its supported hardware to include multi-core CPUs, GPUs, and Field-Programmable Gate Arrays (FPGAs) by designing an OpenCL backend, allowing the incorporation of a wide range of hardware devices. We address OpenABL’s limitation of 2-dimensional and 3-dimensional simulation environments by supporting graph-based simulation spaces in OpenABLext, opening it up for simulation types such as transport simulation which commonly use graph-based networks. A conflict resolution mechanism is introduced to ensure the simulation results of a parallel simulation run do not deviate from a sequential run, significantly improving reproducibility. Eventually, an online dispatcher is proposed as part of OpenABLext to enable automatic best suitable hardware selection for different simulation parts. In a comprehensive simulation study, we show OpenABLext not only drastically simplifies the programming effort but also produces simulation code that outperforms other similar state-of-the-art approaches. With OptCL, we extend the idea of the online dispatcher to further boost the performance of the generated code on top of OpenABLext by enabling a collaborative execution (co-execution) mode where multiple hardware devices work simultaneously on different simulation parts. Through static analysis of data dependencies among compute-intensive code regions and performance predictions, the tool selects the best hardware for each simulation part as well as the execution schemes. The OptCL middleware operates at a language independent layer, a so-called Intermediate Representation. Therefore, it can not only benefit OpenABLext but also many other C-like Domain Specific Languages (DSLs). We demonstrate the power and versatility of OptCL by applying it to OpenABLext as well as another commonly used DSL called SYCL. In both scenarios, up to 21x speed-ups can be achieved. The full framework proposed in this thesis significantly reduces the effort of employing heterogeneous hardware environments for agent-based simulation and thereby paves the way for the high-performance execution for a large variety of simulation models.