Distributed shared memory frameworks: Comparison of implementations

Abstract

A Distributed Shared Memory (DSM) is a memory architecture where separated memories in different nodes can be addressed as one shared address space. It is created by a group of nodes, where each node provides a part of its local memory. Every node of the group is allowed to read or write to the DSM. This enables parallel calculations of complex tasks in multiple nodes and thus increases performance compared to a single-node calculation. DSM frameworks provide mechanisms to create and access a DSM. There are many challenges in the field of DSMs that need to be addressed by the frameworks. These challenges are identified and described in this paper. With regard to these challenges, the DSM frameworks UPC++, DASH, OpenSHMEM, Titanium and Co-Array Fortran are investigated and compared with each other. All investigated frameworks are PGAS implementations. PGAS is one way to implement a DSM, on which this paper is focused. For some challenges the approaches of the frameworks are very similar. However, there are also challenges for which the frameworks offer very different solutions.