Evaluating the running time of a communication round over the internet

Abstract

Idit KeidarThe Technion and MIT idish @ee.technion.ac.il ABSTRACTWe study the running time of distributed algorithms deployed in a widely distributed setting over the Internet using TCP. We consider a simple primitive that corresponds to a communication round in which every host sends information to every other host; this primitive occurs in numerous distributed algorithms. We experiment with four algorithms that typically implement this primitive. We run our experiments on ten hosts at geographically disperse locations over the Internet. We observe that message loss has a large impact on algorithm running times, which causes leader-based algorithms to usually outperform decentralized ones.

Modeling and Validating the Performance of Atomic Broadcast Algorithms in High Latency Networks

“…Bakr and Keidar evaluate the duration of a communication round on the Internet in [3]. Their work focuses on the running time of four distributed algorithms with different message exchange patterns, and in particular, the effect of message loss on these algorithms.…”

“…an algorithm waits until it has received all messages it is expecting). The scope of [3] is similar to ours in that they analyze the relative performance of algorithms with different communication patterns on a wide area network. However, their algorithms are not representative of failure detector based atomic broadcast algorithms.…”

“…Indeed, in the three algorithms we consider, processes never need to wait for messages from all the other processes. Thus, if messages from one process are delayed because of a high-latency link, it does not necessarily affect the performance of the atomic broadcast algorithm (whereas it would in [3]).…”

The mutable consensus protocol

“…With a centralized protocol [7], votes are collected by the coordinator and relayed to other participants only after a decision has been reached, reducing the load on the network at the expense of an additional communication step. Such differences have a definite impact in performance measured in realistic settings [2].…”

RPC-V: Toward Fault-Tolerant RPC for Internet Connected Desktop Grids with Volatile Nodes

“…To implement a fault-tolerant stateful services, servers have to agree on a global order of RPC events which is a form of consensus. In [3], the author shows that agreement protocols, even on pseudo-synchronous networks are very slow. In conjunction with the changes of the system size, agreement may not be meaningful.…”