This report motivates and explains concepts developed in the project MoDiS to organize large virtual address spaces comprising fine-grain concurrent computations in parallel and distributed environments. The single distributed address space is adaptively partitioned by a dynamic set of cooperating managers. The partitioning scheme is decentralized and scales with growing system configurations. Deficiencies as known from centralized or static organizations are prevented. In contrast to common operating systems, thoughts have also been given to possible thread stack and heap overflows and collisions. Both stacks and heaps associated with a thread are realized non-contiguously with linear stack and heap segments to enable the desired exploitation of the possibly large virtual address space. Distribution of data is coupled with garbage collection and based on objects instead of pages while still making use of the hardware faulting mechanism. Implementation is based on ``off the shelf'' hardware components. Crucial for the efficiency of this approach is a thorough top-down oriented construction of all operating system entities comprising the compiler and libraries as well as the kernel.
«