On the Practicality of Atomic MWMR Register Implementations

TL;DR

This paper empirically evaluates four MWMR atomic register algorithms to understand the trade-offs between communication and computation, demonstrating their practicality in real-time distributed systems.

Contribution

It provides an experimental comparison of four MWMR atomic register implementations across simulation and real-world platforms, highlighting their practical performance.

Findings

SFW performs well in terms of latency and communication.

APRX-SFW and CWFR reduce computation at some cost to latency.

The traditional SIMPLE algorithm remains competitive in certain scenarios.

Abstract

Multiple-writer/multiple-reader (MWMR) atomic register implementations provide precise consistency guarantees, in the asynchronous, crash-prone, message passing environment. Fast MWMR atomic register implementations were first introduced in Englert et al. 2009. Fastness is measured in terms of the number of single round read and write operations that does not sacrifice correctness. In Georgiou et al. 2011 was shown, however, that decreasing the communication cost is not enough in these implementations. In particular, considering that the performance is measured in terms of the latency of read and write operations due to both (a) communication delays and (b)local computation, they introduced two new algorithms that traded communication for reducing computation. As computation is still part of the algorithms, someone may wonder: What is the trade-off between communication and local computation in real-time systems? In this work we conduct an experimental performance evaluation of four MWMR atomic register implementations: SFW from Englert et al. 2009, APRX-SFW and CWFR from Georgiou at al. 2011, and the generalization of the traditional algorithm of Attiya et al. 1996 in the MWMR environment, which we call SIMPLE. We implement and evaluate the algorithms on NS2, a single-processor simulator, and on PlanetLab, a planetary-scale real-time network platform. Our comparison provides an empirical answer to the above question and demonstrates the practicality of atomic MWMR register implementations.