DistributedMemoryArchitektur

DistributedMem is a term used in computing to denote architectures and systems that expose a distributed global memory space across multiple networked nodes. In such systems, memory is physically located on different machines, but tools and abstractions allow programs to allocate and access data as if there were a single address space. DistributedMem covers a range of designs, from distributed shared memory DSM attempts that emulate a coherent shared memory to partitioned global address space PGAS approaches that give each node ownership of a portion of memory while providing global access.

Architecture commonly involves multiple compute nodes connected by a network; a memory subsystem that is partitioned; software layers such as runtimes, libraries, and runtime systems that map logical addresses to physical memory; and communication mechanisms (message passing, one-sided remote memory operations). Programming models include MPI with remote memory access (RMA), PGAS languages like UPC, Co-Array Fortran, Chapel, and DSMS implementations with coherence protocols.

Coherence and consistency: DSM strives for a coherent view, often implementing cache coherence protocols and page-based migrations. PGAS prioritizes data locality and explicit remote access semantics; consistency models vary from strict to relaxed, affecting synchronization overhead.

Performance and applications: distributed memory systems aim to scale capacity and throughput for workloads such as high-performance computing simulations, large-scale graph analytics, and data-intensive tasks. Challenges include higher latency for remote accesses, complexity of data placement, and fault tolerance.

See also: distributed shared memory, PGAS, MPI, DSM, NUMA.