Infoscience

Thesis

A multitasking and data-driven architecture for multi-agents simulations

The expansion of 3D real-time simulations (3DRTS) into millions of homes together with the technical progress of computers hardware force to approach software developments for 3DRTS from different perspectives. From an historical standpoint, 3DRTS started principally as homebrew developments. The underlined consequences are the lack of standardization for producing such applications. Nowadays, computers hardware can reproduce close to photo-realism 3D images within interactive environments. This was made possible with the continuous improvements in computers hardware. During many years, the hardware evolution was following vertical speed-up improvements, by increasing CPU clocks speed and memory bandwidth. Today, we are reaching the limits of this approach from a power consumption, heat, and intrinsic materials characteristics perspectives. As an outcome, the next-generation of computer hardware and home consoles are presenting multitasking architectures. This obliges to re-think software development for 3DRTS, moving from the serial and single-threaded approach to a concurrent design. We explore conceptual designs handling the current scale and complexity offered by 3DRTS developments by adopting stronger engineering practices. This is needed to control the underlined complexity and rising developments costs. The direct consequence of being able to generate highly detailed virtual worlds is to involve more deeply artists and designers in the development process. We propose mechanisms that free developers from common low-levels problematic, such as memory management or data synchronization issues. Our architecture relies on extending the Component Based Development (CBD) model for multitasking architectures. This obliges to define specific patterns either directly inspired by other fields in computer science or dedicated for 3DRTS. This includes promoting multi-layer design where the low-level routines are tightly connected to computer hardware by describing the importance of conceiving hardware-oblivious systems. This is important, as memory bandwidth is becoming the principal bottleneck in current applications. Another fundamental aspect consists to move from the single iterative global loop commonly found in single-threaded systems, by incorporating mechanisms for balancing the workflow more accurately. If those optimizations and evolutions are required for assuring efficient real-time performance, they do not allow non-programmers to interact with the system with ease. Our method consists to promote high-level languages and concurrent model relying on Microthreads. This gives the ability to develop and execute scripts in a multitasking environment without the common C/C++ issues. This is primordial to let designers experiment with ideas in a safer and efficient environment. This will leads to adopt the data-driven paradigm to control agents in our simulations, by clearly separating the logic and data layers. This offer better flexibility and reduce the existence of simulation specific code. In addition, we illustrate that the best technology and designs have a limited meaning, if they do not come with a complete production pipeline for managing and controlling simulation assets. This also affects fine tuning parameters where different hardware may perform better in some areas or worse in other. Finally, different use-cases demonstrate the strong and weakness aspects of our approach.

Related material