The Case for Reconfiguration without Consensus: Comparing Algorithms for Atomic Storage

We compare different algorithms for reconfigurable atomic storage in the data-centric model. We present the first experimental evaluation of two recently proposed algorithms for reconfiguration without consensus and compare them to established algorithms for reconfiguration both with and without consensus. Our evaluation reveals that the new algorithms offer a significant improvement in terms of latency and overhead for reconfiguration without consensus. Our evaluation also shows that reconfiguration without consensus, can obtain similar results to that of consensus-based reconfiguration, which relies on a stable leader. Moreover, the new algorithms also substantially reduces the overhead compared to consensus-based reconfiguration without a leader. While our analysis confirms our intuition that batching reconfiguration requests serves to reduce the overhead of reconfigurations, our evaluation also shows that it is equally important to separate reconfigurations from read and write operations. Specifically, we found that using read and write operations to assist in completing concurrent reconfigurations is in fact detrimental to the reconfiguration performance. 1998 ACM Subject Classification C.3.4 Distributed Systems, C.4 Performance of Systems