Towards Efficient Execution of a GALS MoC Based System Level Language

Abstract

Embedded systems include a broad range of electronic systems, from household appliances to safety critical systems such as medical systems, automotive electronics and avionics. Due to growing complexity of the applications, these systems consist of a number of computational intensive units running concurrently. They also interact with each other and with environment, repeatedly reading inputs, doing computations and responding appropriately. These computational units may have different response times; hence, they may need to run concurrently at different speeds. These systems may be considered as GALS (Globally Asynchronous Locally Synchronous), which typically consist of a collection of components that execute concurrently and communicate using possibly slow or unreliable channels. SystemJ is a system level programming language based on GALS model of computation allowing the asynchronous coupling of synchronous reactive modules at the top level, which execute at different speeds. It extends Java with Esterel-like constructs for the synchronous concurrency and reactivity, and CSP-like constructs for the asynchronous concurrency. SystemJ targets a large range of heterogeneous embedded systems that combine data-intensive and control-dominated computations (heterogeneous) in addition to synchronous and asynchronous concurrency. Although, the problem of modeling complex systems has largely been solved by raising level of abstraction, there is still need for efficient execution platforms to realize such applications. While, there have been efforts towards supporting heterogeneous applications, they primarily focused on the reactive part of applications. In summary, developing architectural support for heterogeneous embedded applications has been the main focus of this research. This thesis proposes improvements to some existing architectures as well as developing new architectures that make use of the formal underlying structure of the language to achieve higher execution efficiency in the GALS paradigm. These architectures execute control and data-driven operations along with asynchronous and synchronous concurrent processes in an efficient way. Our novel solutions range from extensions to a single CPU architecture to multiprocessor architectures, all while considering computational demands and resource constraints. We started with the deployment of Java Optimized Processor (JOP) to execute SystemJ programs compiled to Java and improved it by extending its architecture to include the reactive features. We suggested novel architectures which efficiently execute the SystemJ programs compiled by separating control and data-operations on a single core by embedding control operations inside the data-operation represented in Java by translating them to custom bytecodes. We refined this approach by providing two separate modes of execution for control and data-oriented operations. This approach is further extended to multiprocessor architecture to speed up the execution and meet the computational demands required by high-end embedded systems. We experimentally evaluated all proposed architectures to validate their effectiveness. Better performance, lower code density and resources usage have been achieved compared to previous approaches for SystemJ execution, thus proving its suitability for heterogeneous embedded applications.