17. 07. 2015

Implementation of Distributed Shared Memory for Modern Paradigms of Parallel Programming on L4

Matthias Hille

TU Dresden

Verteidigung der Master-Arbeit

Solving problems requiring large compute capacity is frequently done by means of clusters these days. Such clusters consist of many independent compute nodes which are connected by standard networking technology. Writing programs for clusters is more complicated than writing parallel applications for one node due to the lack of shared memory. To remove this development barrier Distributed Shared Memory (DSM) systems have been developed in the past but turned out to be quite inefficient in practice. Nowadays new paradigms in parallel programming such as asynchronous lambdas utilize weaker consistency models, which expose new optimization potential for DSM systems.

To test the feasibility of distributed shared memory using relaxed con- sistency models a DSM infrastructure employing the scope consistency model is built on top of the Fiasco.OC / L4Re environment. The choice of scope consistency facilitates the system to minimize the consistency related network traffic without requiring the programmer to specify hints for every shared variable. With reduced network traffic the DSM system becomes capable of efficiently executing shared memory programs in a distributed environment.

	28. Oct 2020
· Copyright © 2001-2022 Operating Systems Group, TU Dresden \| Impressum ·