Invited Paper: Towards Practical Atomic Distributed Shared Memory: An Experimental Evaluation

Distributed Shared Storage Services may serve as building blocks to yield complex, decentralized, cloud applications in emerging technologies (e.g., IoT, VR/AR), as they offer a transparent cloud storage space where distributed applications can store, retrieve, and coordinate over shared data. Ideally, distributed applications would like to communicate through a "cloud" memory layer that may provide similar guarantees as a centralized sequential memory. Atomic Distributed Shared Memory (ADSM) provides the illusion of a sequential memory space despite asynchrony, network perturbations, and device failures. A plethora of algorithmic solutions along with proven correctness guarantees have been proposed to provide ADSM in a message passing system. None of them, however, has been adopted in a real working solution: commercial solutions avoid the use of ADSM algorithms, mainly due to their communication overhead. But what is exactly the performance overhead of an ADSM algorithm over existing commercial solutions? In this work we want to provide a first answer to this question by performing an in-depth experimental comparison of the state-of-the-art dynamic ADSM algorithm ARES, with two well-established open-source distributed storage solutions, Cassandra and Redis. The results show that ARES's performance is comparable with the commercial systems, with respect to scalability, object size and throughput.