Supporting type-safe languages on DSM systems
Cox, Alan L.
Doctor of Philosophy
In this thesis, we present a new approach to support transparent sharing of data in a distributed system using modern programming languages, such as Java and Modula-3. Unlike RPC-based techniques such as CORBA and RMI, we provide transparent data sharing using software distributed shared memory (DSM). We find that modern programming languages provide new opportunities for optimization while the garbage collection in such languages provides new implementation challenges. This thesis centers around the following two claims. First, the run-time type information and safety features in modern programming languages provide new opportunities to support both coarse-grained and fine-grained sharing efficiently on DSM systems. Our new DSM system, DOSA, takes advantage of the language support to efficiently maintain coherency at the object granularity, making fine-grained sharing efficient. It aggregates the communication of objects to make coarse-grained sharing efficient. Second, the overall performance of garbage collected programs on DSM systems can approach that of manual memory management. Our approach achieves this goal by making the implementation of the garbage collector orthogonal to the DSM operations, and by identifying garbage earlier and more accurately. To substantiate the claims above, we have evaluated our approach against state-of-the-art methods. We first compared the performance of DOSA with that of TreadMarks, a state-of-the-art coarse-grain DSM system. Our evaluation shows that the performance of coarse-grained programs is comparable (within 6%) with TreadMarks, while the performance of fine-grained programs is significantly better (up to 98%) than TreadMarks. We also compared our DSM garbage collector with two state-of-the-art collectors. Our evaluation shows that, with our collector, the overall performance of the programs is close (within 5%) to that of manual management, and far better (up to 50%) than with the existing DSM collectors.