Technical Report Number
Data, Computer Systems Organization, Computing Methodologies, Theory of Computation
This dissertation proposes a new garbage-collected memory module architecture for hard real-time systems. The memory module is designed for compatibility with standard workstation architectures, and cooperates with standard cache consistency protocols. Processes read and write garbage- collected memory in the same manner as standard memory, with identical performance under most conditions. Occasional contention between user processes and the garbage collector results in delays to the user process of at most six memory cycles. Thus the proposed architecture guarantees real-time performance at fine granularity. This dissertation investigates the viability of the proposed architecture in two senses. First, it demonstrates that a fundamental component of the architecture, the object space manager, can be produced at a reasonable cost. Second, this dissertation reports the results of experiments that measure the performance of the proposed architecture under real workloads. Results of these experiments show that the architecture currently performs more slowly than traditional schemes; but this appears to be correctable by employing a more efficient function call mechanism that caches heap- allocated activation frames. Finally, this dissertation reports on some simple extensions to the C++ programming language to support slice objects. Slice objects, which are supported by the garbage collection architecture, are useful for implementing fragmentable arrays, i.e., arrays in which subarrays may be retained while unused elements become garbage and are collected. Experimental evidence demonstrates that slice objects can be used to implement strings more efficiently than at least some popular class libraries.